I built a Retrieval-Augmented Generation chatbot for this portfolio so visitors can ask questions about my research, publications, technical projects, experience, and academic activities without manually navigating across multiple pages.
The retrieval pipeline is fully benchmarked on 100 manually grounded questions, with BM25 selected after reaching 94% Hit@3. The finalized browser demo combines BM25 candidate retrieval, compact evidence selection, and Qwen3-0.6B for grounded local answer generation.
The system answers portfolio questions from controlled evidence while keeping retrieval and model inference in the browser, without a paid API or server-side RAG backend.
Visitors can ask questions about my research, publications, experience, events, and RAG projects directly from the portfolio.
The deployed page loads a compact 88-chunk knowledge base, builds BM25 locally, selects concise evidence, and runs Qwen3-0.6B in the browser. No external inference API or server-side RAG service is required.
A visitor asks a natural-language question directly from the portfolio chatbot.
The browser searches the prepared portfolio corpus and ranks the most relevant chunks.
The strongest retrieved passages become the controlled context for answer generation.
Qwen3-0.6B generates a concise answer from the selected evidence; verified extractive text is used if generation fails.
Retrieved passages retain source metadata internally for debugging and evaluation, while the visitor-facing page shows the exact text rather than internal identifiers.
Selected the portfolio pages, RAG project pages, events, and publication summaries that should form the chatbot knowledge base.
Cleaned the source content and created heading-aware chunks with source, section, category, and traceability metadata.
Embedded every chunk locally with Xenova/all-MiniLM-L6-v2 and stored the normalized 384-dimensional vectors as a compact static asset.
Created 100 manually grounded questions with graded relevance labels so retrieval can be measured independently from answer generation.
Compared Dense MiniLM, BM25, and equal-weight RRF, then inspected which questions and passages each method retrieved or missed.
The homepage presents the project summary, while the complete browser chatbot and technical retrieval analysis remain isolated on this dedicated project page.
BM25 retrieves candidate chunks, the evidence selector keeps up to three compact supporting units, and Qwen3-0.6B rewrites only that controlled context. The model loads lazily on the first generation request and is warmed before producing the full answer.
Run the finalized pipeline over a fixed evaluation set and save each question, retrieved evidence, reference answer, Qwen answer, fallback status, and latency. Evaluate faithfulness/groundedness, factual correctness, answer relevance, and semantic similarity with an automated judge, compare with a stronger secondary judge where useful, and manually review a representative subset before reporting generation-quality results.
This dedicated page contains the complete Edge-RAG demo. Retrieval and evidence selection run locally in the browser, and Qwen is loaded only when a question reaches the generation stage.
Start by loading the lightweight portfolio corpus. Qwen loads only when you ask the first question, after relevant portfolio sentences have been selected.
Local browser experiment. No API key or server is used. Qwen is loaded lazily; verified evidence is used automatically if generation fails.
Dense semantic retrieval remains useful, but on this compact portfolio corpus exact terminology and named entities make lexical matching especially effective.
The main result is shown above. Open this section only if you want to inspect the missed benchmark questions, exact retrieved text, BM25 recovery cases, and retrieval-transition analysis.
These optional details show questions for which no manually labelled relevant passage appeared among the first three retrieved results.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2025 3rd SERENADE Training School The 3rd SERENADE Training School in Padua (22–26 September 2025) combined scientific presentations, industrial visits, and networking activities. It highlighted interdisciplinary research across sensing, AI, food packaging, and industrial applications while strengthening collaboration within the SERENADE network. Padua, Italy
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
The sensing modality changes, but the research logic remains consistent: understand the measurement, preserve informative structure, learn representations that generalise, and design for practical acquisition and deployment. Acquire informative IMU time series or impedance spectra. Expose temporal, spectral, and physical signal structure. Build compact, transferable, and interpretable representations. Reduce sensing, computation, and instrumentation cost. Wearable inertial signals describe human behaviour; impedance spectra describe chemical interactions.
Methods shared across modalities Time-series modelling Representation learning Experimental validation Resource-aware inference Embedded deployment
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
04 · Doctoral flagship Temperature–frequency-tuned impedance spectroscopy Models full complex-impedance spectra across operating temperatures to improve gas discrimination, interpretability, and mixture analysis EIS MOS sensors Open data Research Square preprint ↗
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Wearable and physical signals Human motion through wearable inertial sensors Since 2019, I have studied how body-worn accelerometers and gyroscopes can support reliable human-centred intelligence, spanning activity recognition, gait-based affective computing, cross-domain representation transfer, and sleep monitoring.
Generisch-Net is a generic BiGRU–CNN model for several wearable-IMU human-motion tasks rather than one problem-specific network. The study evaluates the same model for activity recognition, closed-loop person re-identification from gait, and emotion recognition.
This study introduces Jazbat-Net, a lightweight model that transfers representations learned from a large wearable human-activity dataset to emotion recognition from gait. It evaluates smartwatch and smartphone inputs and demonstrates compact, on-device emotion inference.
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Khail-Net is a shallow convolutional network for sports activity recognition from wearable inertial sensors. With 1,251 trainable parameters, it is designed as a resource-conscious model and is evaluated on sports activities as well as a benchmark WISDM human-activity dataset.
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
This study introduces Jazbat-Net, a lightweight model that transfers representations learned from a large wearable human-activity dataset to emotion recognition from gait. It evaluates smartwatch and smartphone inputs and demonstrates compact, on-device emotion inference.
Generisch-Net is a generic BiGRU–CNN model for several wearable-IMU human-motion tasks rather than one problem-specific network. The study evaluates the same model for activity recognition, closed-loop person re-identification from gait, and emotion recognition.
GLiNER receives the 25 reviewed labels and the original chunks. Each stored prediction preserves the exact span, canonical type, confidence, and character offsets. Schema 25 reviewed types Natural-language labels supplied to GLiNER + Text 20 chunks 5 policy documents → Model GLiNER medium v2.1 urchade/gliner_medium-v2.1 → Output 361 detections span + type + score + offsets
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
The benchmark deliberately includes paraphrases, scenarios, thresholds, negative conditions, location rules, approval conditions, historical/current distinctions and cross-policy questions. 64 hard questions 36 medium questions 51 multi-evidence 19 multi-source 100 total questions
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Khail-Net is a shallow convolutional network for sports activity recognition from wearable inertial sensors. With 1,251 trainable parameters, it is designed as a resource-conscious model and is evaluated on sports activities as well as a benchmark WISDM human-activity dataset.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Khail-Net is a shallow convolutional network for sports activity recognition from wearable inertial sensors. With 1,251 trainable parameters, it is designed as a resource-conscious model and is evaluated on sports activities as well as a benchmark WISDM human-activity dataset.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
COUNTRY 38 EMPLOYEE_TYPE 38 POLICY 34 LEAVE_CONCEPT 26 DURATION 25 CURRENCY 22 LIMIT 20 APPROVAL_REQUIREMENT 19 ORGANIZATION 15 BUSINESS_TRAVEL 14 DATE 12 EXPENSE_CONCEPT 12 LEAVE_RULE 11 REMOTE_WORK_ALLOWANCE 11 TRANSPORT_MODE 11 WORK_ARRANGEMENT 11 ENTITLEMENT 9 JOB_ROLE 9 EXCEPTION 7 REQUIREMENT 5 TIME_PERIOD 4 MONEY_AMOUNT 3 COMPANY_RESOURCE 2 TIME_CONSTRAINT 2 HOTEL_RATE_LIMIT 1
September 2024 SERENADE 1st Training School Saarbrücken, Germany
The current GraphRAG work intentionally stops after entity validation. The next research stage will construct relationships between the validated entities before anything is loaded into Neo4j.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
September 2025 3rd SERENADE Training School The 3rd SERENADE Training School in Padua (22–26 September 2025) combined scientific presentations, industrial visits, and networking activities. It highlighted interdisciplinary research across sensing, AI, food packaging, and industrial applications while strengthening collaboration within the SERENADE network. Padua, Italy
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2024 SERENADE 1st Training School Saarbrücken, Germany
May 2025 2nd SERENADE Training School The 2nd SERENADE Training School in Zaragoza (26–30 May 2025) combined theoretical and hands-on training in food science, microbiology, gas sensors, materials, and volatile-compound analysis. The programme also included a visit to BSH Home Appliances, team-building activities, and networking among doctoral candidates and researchers. Zaragoza, Spain
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
Size / overlap Chunks Hit@1 Recall@1 Recall@3 Prec@3 MRR@5 400 / 80 38 0.800 0.725 0.950 0.383 0.892 700 / 120 22 0.800 0.775 1.000 0.367 0.883 1000 / 150 16 0.700 0.650 1.000 0.367 0.825 700 / 120 was selected: complete Recall@3 with fewer chunks than 400/80; 1000/150 reduced first-rank quality. Failure diagnosis: dense retrieval was usually finding the answer, but semantically similar language from unrelated HR policies could enter Top-3. This shifted the research question from “how do I retrieve more?” to “how do I keep the search space cleaner?”
March 2018 Headstart Robotics Competition Islamabad, Pakistan
September 2024 SERENADE 1st Training School Saarbrücken, Germany
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
The benchmark deliberately includes paraphrases, scenarios, thresholds, negative conditions, location rules, approval conditions, historical/current distinctions and cross-policy questions. 64 hard questions 36 medium questions 51 multi-evidence 19 multi-source 100 total questions
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
affective computing, cross-domain representation transfer, and sleep monitoring. Human activity recognition from accelerometer and gyroscope time series Transfer of motion representations to data-scarce emotion tasks Wearable AIoT systems for sleep–wake monitoring Compact CNN, recurrent, and Transformer-based models for edge use IEEE IoT Journal ↗ IEEE Sensors Journal ↗ Wearable motion intelligence Motion signals → Reusable representations → Human-centred tasks
2022 IEEE Sensors Letters Single-author Efficient HAR Khail-Net: A Shallow Convolutional Neural Network for Recognizing Sports Activities Using Wearable Inertial Sensors Explores a shallow, resource-conscious convolutional design for sports-activity recognition from wearable inertial signals. Publisher PDF ↗
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Dense · MRR@3 .850 Routed · MRR@3 .888 Reranked · MRR@3 .917 The cross-encoder improved first-relevant ranking for 11 questions, worsened 7, and left 82 unchanged.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2024 SERENADE 1st Training School Saarbrücken, Germany
August 2026 — Present Research Secondment at BSH Home Appliances Research secondment at BSH Home Appliances in Zaragoza, Spain, started on Monday, 31 August 2026. Zaragoza, Spain
September 2025 Eurosensors 37 Poster presentation on AC excitation-frequency optimisation for MOS gas sensing. Wrocław, Poland
August 2026 — Present Research Secondment at BSH Home Appliances Research secondment at BSH Home Appliances in Zaragoza, Spain, started on Monday, 31 August 2026. Zaragoza, Spain
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
August 2026 — Present Research Secondment at BSH Home Appliances Research secondment at BSH Home Appliances in Zaragoza, Spain, started on Monday, 31 August 2026. Zaragoza, Spain
September 2024 SERENADE 1st Training School Saarbrücken, Germany
September 2025 Eurosensors 37 Poster presentation on AC excitation-frequency optimisation for MOS gas sensing. Wrocław, Poland
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
32 / 100 generated answers changed
Ground truth stores exact source evidence, allowing Hit@K, Recall@K, Precision@K and MRR to be measured without depending on the generator.
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
TF-IDF was used only to select likely policy files. Dense Nomic retrieval still ranked chunks inside that candidate search space. Query User question No manual keyword rules → Route TF-IDF policy scores Up to 2 candidate files; fallback when weak/ambiguous → Embed Nomic query vector search_query: prefix → Retrieve Chroma Top-3 Dense search within selected scope
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
Candidate concepts were collected from the policy corpus, embedded with Nomic, projected with PCA, clustered with HDBSCAN, and manually reviewed into a domain-specific schema. Input HR policies Clean Markdown + 700/120 chunks → Candidates spaCy NER + noun phrases → Represent Nomic embeddings 768-dimensional vectors → Explore PCA + HDBSCAN Structure candidate concepts → Review Manual curation Merge, split, reject noisy clusters → Output 25 entity types Reviewed GLiNER schema
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
Manual review turns the discovered structure into a stable domain schema.
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
Candidate concepts were collected from the policy corpus, embedded with Nomic, projected with PCA, clustered with HDBSCAN, and manually reviewed into a domain-specific schema. Input HR policies Clean Markdown + 700/120 chunks → Candidates spaCy NER + noun phrases → Represent Nomic embeddings 768-dimensional vectors → Explore PCA + HDBSCAN Structure candidate concepts → Review Manual curation Merge, split, reject noisy clusters → Output 25 entity types Reviewed GLiNER schema
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Metric Dense only TF-IDF + dense TF-IDF + dense + reranker Hit@1 0.730 0.790 0.840 Hit@3 0.990 1.000 1.000 Recall@3 0.938 0.938 0.947 Evidence Precision@3 0.420 0.413 0.420 Source Precision@3 0.847 0.950 0.963 MRR@3 0.850 0.888 0.917
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Candidate concepts were collected from the policy corpus, embedded with Nomic, projected with PCA, clustered with HDBSCAN, and manually reviewed into a domain-specific schema. Input HR policies Clean Markdown + 700/120 chunks → Candidates spaCy NER + noun phrases → Represent Nomic embeddings 768-dimensional vectors → Explore PCA + HDBSCAN Structure candidate concepts → Review Manual curation Merge, split, reject noisy clusters → Output 25 entity types Reviewed GLiNER schema
Exact Evidence Precision@3 stayed at 0.367 , because a correct-policy chunk can be useful without containing the exact evidence string. Source Precision@3 exposed the improvement: Dense only ~.750 TF-IDF + dense ~.967
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
COUNTRY 38 EMPLOYEE_TYPE 38 POLICY 34 LEAVE_CONCEPT 26 DURATION 25 CURRENCY 22 LIMIT 20 APPROVAL_REQUIREMENT 19 ORGANIZATION 15 BUSINESS_TRAVEL 14 DATE 12 EXPENSE_CONCEPT 12 LEAVE_RULE 11 REMOTE_WORK_ALLOWANCE 11 TRANSPORT_MODE 11 WORK_ARRANGEMENT 11 ENTITLEMENT 9 JOB_ROLE 9 EXCEPTION 7 REQUIREMENT 5 TIME_PERIOD 4 MONEY_AMOUNT 3 COMPANY_RESOURCE 2 TIME_CONSTRAINT 2 HOTEL_RATE_LIMIT 1
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Wearable accelerometers and gyroscopes · electrical impedance spectroscopy · MOS gas sensors · experimental sensor characterisation Human activity recognition · transfer learning · time-series modelling · compact deep learning · interpretable spectral analysis Python · C/C++ · embedded Linux · Arduino/ESP32 · data pipelines · reproducible research tooling
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
COUNTRY 38 EMPLOYEE_TYPE 38 POLICY 34 LEAVE_CONCEPT 26 DURATION 25 CURRENCY 22 LIMIT 20 APPROVAL_REQUIREMENT 19 ORGANIZATION 15 BUSINESS_TRAVEL 14 DATE 12 EXPENSE_CONCEPT 12 LEAVE_RULE 11 REMOTE_WORK_ALLOWANCE 11 TRANSPORT_MODE 11 WORK_ARRANGEMENT 11 ENTITLEMENT 9 JOB_ROLE 9 EXCEPTION 7 REQUIREMENT 5 TIME_PERIOD 4 MONEY_AMOUNT 3 COMPANY_RESOURCE 2 TIME_CONSTRAINT 2 HOTEL_RATE_LIMIT 1
September 2024 SERENADE 1st Training School Saarbrücken, Germany
The current GraphRAG work intentionally stops after entity validation. The next research stage will construct relationships between the validated entities before anything is loaded into Neo4j.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
32 / 100 generated answers changed
Ground truth stores exact source evidence, allowing Hit@K, Recall@K, Precision@K and MRR to be measured without depending on the generator.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Metric Dense only TF-IDF + dense TF-IDF + dense + reranker Hit@1 0.730 0.790 0.840 Hit@3 0.990 1.000 1.000 Recall@3 0.938 0.938 0.947 Evidence Precision@3 0.420 0.413 0.420 Source Precision@3 0.847 0.950 0.963 MRR@3 0.850 0.888 0.917
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The standalone viewer contains the entire chunked corpus and allows filtering by source document, entity type, and confidence. It is used to check coverage, span boundaries, label quality, and the effect of moving the confidence threshold before relations are trusted.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The standalone viewer contains the entire chunked corpus and allows filtering by source document, entity type, and confidence. It is used to check coverage, span boundaries, label quality, and the effect of moving the confidence threshold before relations are trusted.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2026 Research Square Preprint Gas sensing Temperature–frequency-tuned impedance spectroscopy with interpretable full-spectrum modelling enables robust gas sensing Introduces a temperature- and frequency-resolved framework for compact, interpretable modelling of MOS-sensor impedance in single gases and mixtures. ↗
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2026 Research Square Preprint Gas sensing Temperature–frequency-tuned impedance spectroscopy with interpretable full-spectrum modelling enables robust gas sensing Introduces a temperature- and frequency-resolved framework for compact, interpretable modelling of MOS-sensor impedance in single gases and mixtures. ↗
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2025 3rd SERENADE Training School The 3rd SERENADE Training School in Padua (22–26 September 2025) combined scientific presentations, industrial visits, and networking activities. It highlighted interdisciplinary research across sensing, AI, food packaging, and industrial applications while strengthening collaboration within the SERENADE network. Padua, Italy
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
The sensing modality changes, but the research logic remains consistent: understand the measurement, preserve informative structure, learn representations that generalise, and design for practical acquisition and deployment. Acquire informative IMU time series or impedance spectra. Expose temporal, spectral, and physical signal structure. Build compact, transferable, and interpretable representations. Reduce sensing, computation, and instrumentation cost. Wearable inertial signals describe human behaviour; impedance spectra describe chemical interactions.
Methods shared across modalities Time-series modelling Representation learning Experimental validation Resource-aware inference Embedded deployment
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
COUNTRY 38 EMPLOYEE_TYPE 38 POLICY 34 LEAVE_CONCEPT 26 DURATION 25 CURRENCY 22 LIMIT 20 APPROVAL_REQUIREMENT 19 ORGANIZATION 15 BUSINESS_TRAVEL 14 DATE 12 EXPENSE_CONCEPT 12 LEAVE_RULE 11 REMOTE_WORK_ALLOWANCE 11 TRANSPORT_MODE 11 WORK_ARRANGEMENT 11 ENTITLEMENT 9 JOB_ROLE 9 EXCEPTION 7 REQUIREMENT 5 TIME_PERIOD 4 MONEY_AMOUNT 3 COMPANY_RESOURCE 2 TIME_CONSTRAINT 2 HOTEL_RATE_LIMIT 1
September 2024 SERENADE 1st Training School Saarbrücken, Germany
The current GraphRAG work intentionally stops after entity validation. The next research stage will construct relationships between the validated entities before anything is loaded into Neo4j.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
The benchmark deliberately includes paraphrases, scenarios, thresholds, negative conditions, location rules, approval conditions, historical/current distinctions and cross-policy questions. 64 hard questions 36 medium questions 51 multi-evidence 19 multi-source 100 total questions
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
affective computing, cross-domain representation transfer, and sleep monitoring. Human activity recognition from accelerometer and gyroscope time series Transfer of motion representations to data-scarce emotion tasks Wearable AIoT systems for sleep–wake monitoring Compact CNN, recurrent, and Transformer-based models for edge use IEEE IoT Journal ↗ IEEE Sensors Journal ↗ Wearable motion intelligence Motion signals → Reusable representations → Human-centred tasks
2022 IEEE Sensors Letters Single-author Efficient HAR Khail-Net: A Shallow Convolutional Neural Network for Recognizing Sports Activities Using Wearable Inertial Sensors Explores a shallow, resource-conscious convolutional design for sports-activity recognition from wearable inertial signals. Publisher PDF ↗
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Dense · MRR@3 .850 Routed · MRR@3 .888 Reranked · MRR@3 .917 The cross-encoder improved first-relevant ranking for 11 questions, worsened 7, and left 82 unchanged.
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
32 / 100 generated answers changed
Ground truth stores exact source evidence, allowing Hit@K, Recall@K, Precision@K and MRR to be measured without depending on the generator.
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
TF-IDF was used only to select likely policy files. Dense Nomic retrieval still ranked chunks inside that candidate search space. Query User question No manual keyword rules → Route TF-IDF policy scores Up to 2 candidate files; fallback when weak/ambiguous → Embed Nomic query vector search_query: prefix → Retrieve Chroma Top-3 Dense search within selected scope
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Metric Dense only TF-IDF + dense TF-IDF + dense + reranker Hit@1 0.730 0.790 0.840 Hit@3 0.990 1.000 1.000 Recall@3 0.938 0.938 0.947 Evidence Precision@3 0.420 0.413 0.420 Source Precision@3 0.847 0.950 0.963 MRR@3 0.850 0.888 0.917
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
BM25 did not simply improve the aggregate score. It retrieved manually labelled relevant evidence in the Top-3 for 16 questions where Dense MiniLM returned no relevant chunk.
Dense MiniLM succeeds on 81 of 100 questions. BM25 keeps 78 of those successes, recovers 16 Dense misses, and loses 3 Dense successes, resulting in 94 successful Top-3 retrievals.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
The sensing modality changes, but the research logic remains consistent: understand the measurement, preserve informative structure, learn representations that generalise, and design for practical acquisition and deployment. Acquire informative IMU time series or impedance spectra. Expose temporal, spectral, and physical signal structure. Build compact, transferable, and interpretable representations. Reduce sensing, computation, and instrumentation cost. Wearable inertial signals describe human behaviour; impedance spectra describe chemical interactions.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
This study introduces Jazbat-Net, a lightweight model that transfers representations learned from a large wearable human-activity dataset to emotion recognition from gait. It evaluates smartwatch and smartphone inputs and demonstrates compact, on-device emotion inference.
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
September 2025 Eurosensors 37 Poster presentation on AC excitation-frequency optimisation for MOS gas sensing. Wrocław, Poland
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
This comparison is more informative than aggregate scores alone because it shows whether the two retrieval signals are genuinely complementary.
Lexical retrieval recovered relevant Top-3 evidence for these questions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2025 3rd SERENADE Training School The 3rd SERENADE Training School in Padua (22–26 September 2025) combined scientific presentations, industrial visits, and networking activities. It highlighted interdisciplinary research across sensing, AI, food packaging, and industrial applications while strengthening collaboration within the SERENADE network. Padua, Italy
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
The sensing modality changes, but the research logic remains consistent: understand the measurement, preserve informative structure, learn representations that generalise, and design for practical acquisition and deployment. Acquire informative IMU time series or impedance spectra. Expose temporal, spectral, and physical signal structure. Build compact, transferable, and interpretable representations. Reduce sensing, computation, and instrumentation cost. Wearable inertial signals describe human behaviour; impedance spectra describe chemical interactions.
Methods shared across modalities Time-series modelling Representation learning Experimental validation Resource-aware inference Embedded deployment
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
04 · Doctoral flagship Temperature–frequency-tuned impedance spectroscopy Models full complex-impedance spectra across operating temperatures to improve gas discrimination, interpretability, and mixture analysis EIS MOS sensors Open data Research Square preprint ↗
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Wearable and physical signals Human motion through wearable inertial sensors Since 2019, I have studied how body-worn accelerometers and gyroscopes can support reliable human-centred intelligence, spanning activity recognition, gait-based affective computing, cross-domain representation transfer, and sleep monitoring.
Generisch-Net is a generic BiGRU–CNN model for several wearable-IMU human-motion tasks rather than one problem-specific network. The study evaluates the same model for activity recognition, closed-loop person re-identification from gait, and emotion recognition.
This study introduces Jazbat-Net, a lightweight model that transfers representations learned from a large wearable human-activity dataset to emotion recognition from gait. It evaluates smartwatch and smartphone inputs and demonstrates compact, on-device emotion inference.
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Khail-Net is a shallow convolutional network for sports activity recognition from wearable inertial sensors. With 1,251 trainable parameters, it is designed as a resource-conscious model and is evaluated on sports activities as well as a benchmark WISDM human-activity dataset.
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
This study introduces Jazbat-Net, a lightweight model that transfers representations learned from a large wearable human-activity dataset to emotion recognition from gait. It evaluates smartwatch and smartphone inputs and demonstrates compact, on-device emotion inference.
Generisch-Net is a generic BiGRU–CNN model for several wearable-IMU human-motion tasks rather than one problem-specific network. The study evaluates the same model for activity recognition, closed-loop person re-identification from gait, and emotion recognition.
GLiNER receives the 25 reviewed labels and the original chunks. Each stored prediction preserves the exact span, canonical type, confidence, and character offsets. Schema 25 reviewed types Natural-language labels supplied to GLiNER + Text 20 chunks 5 policy documents → Model GLiNER medium v2.1 urchade/gliner_medium-v2.1 → Output 361 detections span + type + score + offsets
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
The benchmark deliberately includes paraphrases, scenarios, thresholds, negative conditions, location rules, approval conditions, historical/current distinctions and cross-policy questions. 64 hard questions 36 medium questions 51 multi-evidence 19 multi-source 100 total questions
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Khail-Net is a shallow convolutional network for sports activity recognition from wearable inertial sensors. With 1,251 trainable parameters, it is designed as a resource-conscious model and is evaluated on sports activities as well as a benchmark WISDM human-activity dataset.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Spak-Net is a lightweight convolutional network for sports and general activity recognition from wearable inertial sensors. Its addition-based architecture has only 398 trainable parameters and is evaluated on the IM-Sporting Behavior and WISDM-11 datasets with edge deployment in mind.
Khail-Net is a shallow convolutional network for sports activity recognition from wearable inertial sensors. With 1,251 trainable parameters, it is designed as a resource-conscious model and is evaluated on sports activities as well as a benchmark WISDM human-activity dataset.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
September 2025 3rd SERENADE Training School The 3rd SERENADE Training School in Padua (22–26 September 2025) combined scientific presentations, industrial visits, and networking activities. It highlighted interdisciplinary research across sensing, AI, food packaging, and industrial applications while strengthening collaboration within the SERENADE network. Padua, Italy
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2024 SERENADE 1st Training School Saarbrücken, Germany
May 2025 2nd SERENADE Training School The 2nd SERENADE Training School in Zaragoza (26–30 May 2025) combined theoretical and hands-on training in food science, microbiology, gas sensors, materials, and volatile-compound analysis. The programme also included a visit to BSH Home Appliances, team-building activities, and networking among doctoral candidates and researchers. Zaragoza, Spain
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
Size / overlap Chunks Hit@1 Recall@1 Recall@3 Prec@3 MRR@5 400 / 80 38 0.800 0.725 0.950 0.383 0.892 700 / 120 22 0.800 0.775 1.000 0.367 0.883 1000 / 150 16 0.700 0.650 1.000 0.367 0.825 700 / 120 was selected: complete Recall@3 with fewer chunks than 400/80; 1000/150 reduced first-rank quality. Failure diagnosis: dense retrieval was usually finding the answer, but semantically similar language from unrelated HR policies could enter Top-3. This shifted the research question from “how do I retrieve more?” to “how do I keep the search space cleaner?”
March 2018 Headstart Robotics Competition Islamabad, Pakistan
September 2024 SERENADE 1st Training School Saarbrücken, Germany
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
The benchmark deliberately includes paraphrases, scenarios, thresholds, negative conditions, location rules, approval conditions, historical/current distinctions and cross-policy questions. 64 hard questions 36 medium questions 51 multi-evidence 19 multi-source 100 total questions
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
affective computing, cross-domain representation transfer, and sleep monitoring. Human activity recognition from accelerometer and gyroscope time series Transfer of motion representations to data-scarce emotion tasks Wearable AIoT systems for sleep–wake monitoring Compact CNN, recurrent, and Transformer-based models for edge use IEEE IoT Journal ↗ IEEE Sensors Journal ↗ Wearable motion intelligence Motion signals → Reusable representations → Human-centred tasks
2022 IEEE Sensors Letters Single-author Efficient HAR Khail-Net: A Shallow Convolutional Neural Network for Recognizing Sports Activities Using Wearable Inertial Sensors Explores a shallow, resource-conscious convolutional design for sports-activity recognition from wearable inertial signals. Publisher PDF ↗
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Dense · MRR@3 .850 Routed · MRR@3 .888 Reranked · MRR@3 .917 The cross-encoder improved first-relevant ranking for 11 questions, worsened 7, and left 82 unchanged.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2024 SERENADE 1st Training School Saarbrücken, Germany
August 2026 — Present Research Secondment at BSH Home Appliances Research secondment at BSH Home Appliances in Zaragoza, Spain, started on Monday, 31 August 2026. Zaragoza, Spain
September 2025 Eurosensors 37 Poster presentation on AC excitation-frequency optimisation for MOS gas sensing. Wrocław, Poland
August 2026 — Present Research Secondment at BSH Home Appliances Research secondment at BSH Home Appliances in Zaragoza, Spain, started on Monday, 31 August 2026. Zaragoza, Spain
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
August 2026 — Present Research Secondment at BSH Home Appliances Research secondment at BSH Home Appliances in Zaragoza, Spain, started on Monday, 31 August 2026. Zaragoza, Spain
September 2024 SERENADE 1st Training School Saarbrücken, Germany
September 2025 Eurosensors 37 Poster presentation on AC excitation-frequency optimisation for MOS gas sensing. Wrocław, Poland
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
TF-IDF was used only to select likely policy files. Dense Nomic retrieval still ranked chunks inside that candidate search space. Query User question No manual keyword rules → Route TF-IDF policy scores Up to 2 candidate files; fallback when weak/ambiguous → Embed Nomic query vector search_query: prefix → Retrieve Chroma Top-3 Dense search within selected scope
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
Candidate concepts were collected from the policy corpus, embedded with Nomic, projected with PCA, clustered with HDBSCAN, and manually reviewed into a domain-specific schema. Input HR policies Clean Markdown + 700/120 chunks → Candidates spaCy NER + noun phrases → Represent Nomic embeddings 768-dimensional vectors → Explore PCA + HDBSCAN Structure candidate concepts → Review Manual curation Merge, split, reject noisy clusters → Output 25 entity types Reviewed GLiNER schema
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
Manual review turns the discovered structure into a stable domain schema.
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
Candidate concepts were collected from the policy corpus, embedded with Nomic, projected with PCA, clustered with HDBSCAN, and manually reviewed into a domain-specific schema. Input HR policies Clean Markdown + 700/120 chunks → Candidates spaCy NER + noun phrases → Represent Nomic embeddings 768-dimensional vectors → Explore PCA + HDBSCAN Structure candidate concepts → Review Manual curation Merge, split, reject noisy clusters → Output 25 entity types Reviewed GLiNER schema
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The graph is treated as an engineered representation, not as an automatic LLM by-product. Each intermediate stage is inspected before the next one is trusted. 08.1 Schema discovery spaCy candidates → embeddings → PCA/HDBSCAN → manual review → 08.2 Entity extraction GLiNER spans + type + confidence + offsets → 08.2.1 Entity validation Interactive full-corpus inspection → Next · 08.3 Relation extraction Build relations only after entity quality is validated → Future Neo4j + GraphRAG Validated graph retrieval compared with the classical RAG baseline
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Candidate concepts were collected from the policy corpus, embedded with Nomic, projected with PCA, clustered with HDBSCAN, and manually reviewed into a domain-specific schema. Input HR policies Clean Markdown + 700/120 chunks → Candidates spaCy NER + noun phrases → Represent Nomic embeddings 768-dimensional vectors → Explore PCA + HDBSCAN Structure candidate concepts → Review Manual curation Merge, split, reject noisy clusters → Output 25 entity types Reviewed GLiNER schema
Exact Evidence Precision@3 stayed at 0.367 , because a correct-policy chunk can be useful without containing the exact evidence string. Source Precision@3 exposed the improvement: Dense only ~.750 TF-IDF + dense ~.967
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
COUNTRY 38 EMPLOYEE_TYPE 38 POLICY 34 LEAVE_CONCEPT 26 DURATION 25 CURRENCY 22 LIMIT 20 APPROVAL_REQUIREMENT 19 ORGANIZATION 15 BUSINESS_TRAVEL 14 DATE 12 EXPENSE_CONCEPT 12 LEAVE_RULE 11 REMOTE_WORK_ALLOWANCE 11 TRANSPORT_MODE 11 WORK_ARRANGEMENT 11 ENTITLEMENT 9 JOB_ROLE 9 EXCEPTION 7 REQUIREMENT 5 TIME_PERIOD 4 MONEY_AMOUNT 3 COMPANY_RESOURCE 2 TIME_CONSTRAINT 2 HOTEL_RATE_LIMIT 1
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
Dense semantic evidence was genuinely complementary for these cases.
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Wearable accelerometers and gyroscopes · electrical impedance spectroscopy · MOS gas sensors · experimental sensor characterisation Human activity recognition · transfer learning · time-series modelling · compact deep learning · interpretable spectral analysis Python · C/C++ · embedded Linux · Arduino/ESP32 · data pipelines · reproducible research tooling
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
Completed degrees MS, Computer Science National University of Sciences and Technology - School of Electrical Engineering and Computer Science · Islamabad, Pakistan BS, Electrical Engineering National University of Computer and Emerging Sciences · Islamabad, Pakistan
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
I welcome research collaborations and conversations about postdoctoral or industrial R&D roles.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The standalone viewer contains the entire chunked corpus and allows filtering by source document, entity type, and confidence. It is used to check coverage, span boundaries, label quality, and the effect of moving the confidence threshold before relations are trusted.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Current Validated entities Finalize confidence threshold, spans and entity types → 08.3 Relation extraction Generate candidate relations between valid entity pairs → Validate Relation quality Inspect semantics, direction and consistency → Then Neo4j Load only validated graph structure → Evaluate GraphRAG Compare retrieval against the traditional RAG baseline Current boundary of this page: 08.1 schema discovery, 08.2 GLiNER entity extraction and 08.2.1 interactive entity validation are documented here. Relation extraction results will be added only after 08.3 is completed and validated.
The standalone viewer contains the entire chunked corpus and allows filtering by source document, entity type, and confidence. It is used to check coverage, span boundaries, label quality, and the effect of moving the confidence threshold before relations are trusted.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2026 Research Square Preprint Gas sensing Temperature–frequency-tuned impedance spectroscopy with interpretable full-spectrum modelling enables robust gas sensing Introduces a temperature- and frequency-resolved framework for compact, interpretable modelling of MOS-sensor impedance in single gases and mixtures. ↗
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
My academic work applies a consistent sensor-intelligence approach across wearable inertial signals and impedance-based gas sensing, grounded in more than five years of embedded, automation, and machine-learning engineering experience.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2026 Research Square Preprint Gas sensing Temperature–frequency-tuned impedance spectroscopy with interpretable full-spectrum modelling enables robust gas sensing Introduces a temperature- and frequency-resolved framework for compact, interpretable modelling of MOS-sensor impedance in single gases and mixtures. ↗
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Equal fusion pushed relevant BM25 evidence outside the Top-3.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
September 2025 3rd SERENADE Training School The 3rd SERENADE Training School in Padua (22–26 September 2025) combined scientific presentations, industrial visits, and networking activities. It highlighted interdisciplinary research across sensing, AI, food packaging, and industrial applications while strengthening collaboration within the SERENADE network. Padua, Italy
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
The sensing modality changes, but the research logic remains consistent: understand the measurement, preserve informative structure, learn representations that generalise, and design for practical acquisition and deployment. Acquire informative IMU time series or impedance spectra. Expose temporal, spectral, and physical signal structure. Build compact, transferable, and interpretable representations. Reduce sensing, computation, and instrumentation cost. Wearable inertial signals describe human behaviour; impedance spectra describe chemical interactions.
Methods shared across modalities Time-series modelling Representation learning Experimental validation Resource-aware inference Embedded deployment
June 2019 IGNITE National Grassroots Research Initiative Islamabad, Pakistan
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
I am a Marie Skłodowska-Curie doctoral researcher at Saarland University. I work on the broader problem of extracting reliable and deployable intelligence from physical sensor signals, with current doctoral research focused on excitation design, frequency-domain impedance information, and model-based representations for MOS gas sensors. Wearable inertial sensing forms the other major body of this work, developed through human activity recognition, gait-based emotion detection, cross-domain transfer learning, sleep monitoring, and compact edge-oriented neural models.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Recent activities focus on gas sensing, electrical impedance spectroscopy, food-freshness monitoring, research secondments, and interdisciplinary doctoral training.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
2024 — present MSCA Doctoral Researcher Saarland University · SERENADE Doctoral research in Systems Engineering on excitation strategies, impedance modelling, Unsupervised learning, and data visualisations for MOS gas sensors. The programme includes industrial research periods with Bosch Sensortec. Host: Lab for Measurement Technology, Saarland University Industry collaboration: Bosch Sensortec GmbH
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
The benchmark deliberately includes paraphrases, scenarios, thresholds, negative conditions, location rules, approval conditions, historical/current distinctions and cross-policy questions. 64 hard questions 36 medium questions 51 multi-evidence 19 multi-source 100 total questions
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
travel_policy.md 94 remote_work.md 74 annual_leave.md 73 parental_leave.md 61 expense_policy.md 59 Counts are mention-level detections in the current ≥0.20 JSON and include repeated entities across chunks/context.
My work spans wearable inertial sensors and impedance-based chemical sensors, connected by one methodology: design informative measurements, learn compact representations, validate under real conditions, and reduce the cost of deployment. Open to postdoctoral and industrial research opportunities in intelligent sensing and applied machine learning. Based in Germany Saarland University · Bosch Sensortec collaboration Measure Model & learn Deploy 2019+ continuous research in sensor AI 7 journal articles across sensor AI 6 first-author journal articles 5+ years in product engineering
affective computing, cross-domain representation transfer, and sleep monitoring. Human activity recognition from accelerometer and gyroscope time series Transfer of motion representations to data-scarce emotion tasks Wearable AIoT systems for sleep–wake monitoring Compact CNN, recurrent, and Transformer-based models for edge use IEEE IoT Journal ↗ IEEE Sensors Journal ↗ Wearable motion intelligence Motion signals → Reusable representations → Human-centred tasks
2022 IEEE Sensors Letters Single-author Efficient HAR Khail-Net: A Shallow Convolutional Neural Network for Recognizing Sports Activities Using Wearable Inertial Sensors Explores a shallow, resource-conscious convolutional design for sports-activity recognition from wearable inertial signals. Publisher PDF ↗
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Dense · MRR@3 .850 Routed · MRR@3 .888 Reranked · MRR@3 .917 The cross-encoder improved first-relevant ranking for 11 questions, worsened 7, and left 82 unchanged.
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
5 representative questions reviewed manually GPT-5.6 Sol aligned much better with the manual judgments than the smaller local judge in this small spot-check. Interpretation: generation scores are diagnostic, not absolute truth. The disagreement between judges became a result in itself: evaluator quality can become an evaluation bottleneck.
The experiments were intentionally staged: establish a dense baseline, inspect its failures, reduce policy-space noise, test reranking, and only then evaluate answer generation. 01 HR corpus 5 fictional Markdown policies → 02 Chunking 400/80, 700/120, 1000/150 → 03 Nomic + Chroma Dense semantic retrieval → 04 Failure analysis Exact evidence + source relevance → 05 TF-IDF routing Select likely policy file(s) → 06 Optional reranker Cross-encoder Top-5 → Top-3 → 07 Qwen generation Grounded Top-3 context → 08 Evaluation RAGAS + secondary judge + manual check
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
TF-IDF was used only to select likely policy files. Dense Nomic retrieval still ranked chunks inside that candidate search space. Query User question No manual keyword rules → Route TF-IDF policy scores Up to 2 candidate files; fallback when weak/ambiguous → Embed Nomic query vector search_query: prefix → Retrieve Chroma Top-3 Dense search within selected scope
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The first 20-question benchmark showed that Top-3 coverage was already strong, but additional chunks often came from the wrong policy.
Selected conferences, doctoral-network training schools, research secondments, technical courses, poster presentations, and earlier engineering events. 2018—2026 Selected academic and engineering events across Germany, Italy, Poland, Spain, and Pakistan.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
automation, supported technical hiring, and trained engineers in Python, Linux, and Git. Design Engineer (Embedded Systems) Dec 2018 — Dec 2022 Developed C/C++ firmware, custom embedded-Linux and Yocto images, and rapid IoT prototypes for MCUs and single-board computers.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
MONEY_AMOUNT DURATION ORGANIZATION LEAVE_RULE APPROVAL_REQUIREMENT LEAVE_CONCEPT POLICY WORK_ARRANGEMENT REMOTE_WORK_ALLOWANCE BUSINESS_TRAVEL DATE TIME_PERIOD TIME_CONSTRAINT LIMIT HOTEL_RATE_LIMIT EXPENSE_CONCEPT JOB_ROLE ENTITLEMENT REQUIREMENT EMPLOYEE_TYPE COMPANY_RESOURCE EXCEPTION TRANSPORT_MODE COUNTRY CURRENCY
COUNTRY 38 EMPLOYEE_TYPE 38 POLICY 34 LEAVE_CONCEPT 26 DURATION 25 CURRENCY 22 LIMIT 20 APPROVAL_REQUIREMENT 19 ORGANIZATION 15 BUSINESS_TRAVEL 14 DATE 12 EXPENSE_CONCEPT 12 LEAVE_RULE 11 REMOTE_WORK_ALLOWANCE 11 TRANSPORT_MODE 11 WORK_ARRANGEMENT 11 ENTITLEMENT 9 JOB_ROLE 9 EXCEPTION 7 REQUIREMENT 5 TIME_PERIOD 4 MONEY_AMOUNT 3 COMPANY_RESOURCE 2 TIME_CONSTRAINT 2 HOTEL_RATE_LIMIT 1
September 2024 SERENADE 1st Training School Saarbrücken, Germany
The current GraphRAG work intentionally stops after entity validation. The next research stage will construct relationships between the validated entities before anything is loaded into Neo4j.
A selected archive spanning doctoral research, conference participation, funded student engineering, and robotics competitions.
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
The generator was deliberately kept small so that retrieval and grounding quality remained visible rather than being masked by a large model's prior knowledge. Route TF-IDF Policy candidate selection → Retrieve Nomic + Chroma Top-3 context → Generate qwen3:0.6b Context-only HR answer → Judge A RAGAS + Qwen2.5 Local, reproducible → Judge B GPT-5.6 Sol Same saved answers; secondary evaluation
Pipeline Judge Faithfulness Factual correctness Semantic similarity Answer relevancy Dense RAGAS + Qwen2.5 0.801 0.652 0.779 — TF-IDF + Dense RAGAS + Qwen2.5 0.754 0.668 0.778 — Dense GPT-5.6 Sol 0.831 0.753 — 0.912 TF-IDF + Dense GPT-5.6 Sol 0.848 0.775 — 0.921
01 · Completed baseline Traditional RAG FakeCompanyABC: evaluated retrieval before generation A fully local RAG pipeline over fictional HR policies, developed through chunking experiments, dense retrieval, policy routing, reranking, grounded generation, and evaluator comparison. 100 benchmark questions 1.000 Hit@3 ~67 ms selected CPU retrieval Nomic embeddings ChromaDB TF-IDF routing Qwen View project details GitHub ↗
The project showed that retrieval should be diagnosed before adding heavier components. TF-IDF policy routing substantially cleaned the search space at negligible cost; cross-encoder reranking improved ranking further but added ~400 ms CPU latency. The resulting classical pipeline is preserved as the baseline for the GraphRAG study.
32 / 100 generated answers changed
Embeddings: nomic-embed-text (768-D) Vector DB: ChromaDB Generator: qwen3:0.6b Local judge: qwen2.5:1.5b-instruct + RAGAS
32 / 100 generated answers changed
Ground truth stores exact source evidence, allowing Hit@K, Recall@K, Precision@K and MRR to be measured without depending on the generator.
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Metric Dense only TF-IDF + dense TF-IDF + dense + reranker Hit@1 0.730 0.790 0.840 Hit@3 0.990 1.000 1.000 Recall@3 0.938 0.938 0.947 Evidence Precision@3 0.420 0.413 0.420 Source Precision@3 0.847 0.950 0.963 MRR@3 0.850 0.888 0.917
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
02 · Work in progress GraphRAG FakeCompanyABC: constructing a validated policy graph An extension of the same corpus toward GraphRAG, currently focused on entity-schema discovery, GLiNER extraction, and human validation before relation extraction and Neo4j evaluation. 25 entity types 361 stored detections 198 unique entity/type pairs GLiNER Entity validation Neo4j Graph retrieval View project details Open entity viewer ↗
212 detections at old ≥ 0.50 run 98 unique entity/type pairs → 361 detections at current ≥ 0.20 run 198 unique entity/type pairs The 0.20 threshold is an extraction floor for validation , not the final acceptance threshold. Lower-confidence candidates are intentionally retained so they can be inspected instead of discarded upstream.
Documents 5 Chunks 20 Chunk size / overlap 700 / 120 Schema size 25 types Stored detections 361 Unique entity/type pairs 198 Extraction floor 0.20 Confidence range 0.202 – 0.987 Mean confidence 0.586