I build and evaluate AI systems for healthcare and biometrics, with a focus on outputs that can be traced, audited and honestly reported. Long term, I want to work on trustworthy medical AI for neurological disease, including neural-signal decoding to help non-speaking people communicate.
| Challenge / work | Result | My role | Paper | Code | |
|---|---|---|---|---|---|
| 🥇 | StepUP 2026 · footstep biometrics (IJCB) | 1st place · EER 8.00% | Second author | arXiv | repo |
| 🥉 | IJCB-AFMFR 2026 · face recognition, synthetic data | 3rd, Limited Data Track | Co-author | arXiv | repo |
| 📄 | ICICCT 2025 · blockchain complaint system | Published, Springer LNNS 1664 | First author | DOI | |
| 🏥 | Clinical LLM pipeline · discharge summaries | Validated on a 669-test suite | Engineer | confidential | |
| 🌐 | HL7 AI Challenge 2026 · cross-border FHIR IPS | 0 validator errors on 4 IPS routes | Team member | repo |
Plus co-authorship on four more IJCB 2026 competition papers: V-MAD, SSBC, UERC and LivDet-Face (full list).
| Area | Topics |
|---|---|
| Healthcare AI | Clinical information extraction, terminology grounding (SNOMED CT, UMLS, MedDRA), evidence traceability, medical AI evaluation |
| Biometrics | Footstep recognition, face and ear recognition, sclera segmentation, morphing-attack detection |
| Interoperability & multimodal ML | FHIR R4, structured clinical records, speech and audio models |
Spatiotemporal CNN with score fusion for paired pressure footsteps. 8.00% EER on the competition evaluation. Paper (arXiv:2607.13905) · Code Competition data and model checkpoints are not included in the repository.
Llama 3.1 70B with chain-of-thought and few-shot prompting to draft clinical discharge summaries for a healthcare client, validated on a 669-test suite. Stack: GCP, Whisper, MedGemma, SNOMED CT / UMLS / MedDRA. Client code, data and identifying details are confidential.
Reference-constrained transformation of clinical-conversation information into auditable evidence for a frozen population Bayesian network. The mapping does not discover causal structure. Research code is private.
Team prototype that turns synthetic clinical reports into terminology-mapped FHIR R4 International Patient Summary bundles with supporting evidence. Evaluated on synthetic data only, not clinically validated. Repository
Mosquito-species recognition across recording domains. I reviewed the literature, designed the spectrogram structure, developed the approach and contributed code. Reported baseline figures are not a final team score. Repository
Research prototype for report reading and multilingual voice interaction. Uses external services and is not clinically validated; the repository documents its privacy limitations. Repository
7 papers: 1 published (Springer) and 6 IJCB 2026 competition papers. Each is listed at its current status.
Published
- Analytical Approach of False Products and Complaint System Development Using Blockchain Technology. Gupta, A., Pavan Sai, A. and Dubey, S.K. ICICCT 2025, Lecture Notes in Networks and Systems vol. 1664, Springer, 2026. DOI · Springer
Accepted at IJCB 2026 (preprints available)
- The 2nd International StepUP Competition for Biometric Footstep Recognition: From Steps to Strides. Second author. 🥇 Team placed 1st, best EER 8.00%. arXiv:2607.13905 · Code
- Competition on Adapting Foundation Models for Face Recognition Using Synthetic Training Data (IJCB-AFMFR 2026). Co-author (competition participant). 🥉 Team placed 3rd in the Limited Data Track. arXiv:2607.24422 · PDF · Code
IJCB 2026 competition papers (presented; final proceedings pending)
- Every Frame You Take: The Video-based Morphing Attack Detection Challenge (V-MAD). Co-author. Quality-aware dual-branch late fusion: EfficientNet-B0 artifact branch plus ArcFace identity branch with OFIQ frame filtering. PDF
- Label-Efficient Sclera Segmentation with Foundation Models: SSBC 2026. Co-author. SAM-LoRA and DINOv2 team entries. PDF · Code
- The Unconstrained Ear Recognition Challenge 2026: Size Matters (UERC). Co-author. ConvNeXt + ArcFace entry. PDF · Code
- LivDet-Face 2026: Face Liveness Detection Competition. Co-author (team CosmicFace). Competition report
In preparation
- MediLipi: reference-constrained clinical evidence mapping for a frozen population Bayesian network.
B.Tech., Computer Science and Engineering, Amity University Uttar Pradesh (Noida), 2025 AI Research Engineer, ArogyaPandit Private Limited, Gurugram
Python PyTorch LLMs Whisper Google Cloud FHIR R4 SNOMED CT UMLS MedDRA Bayesian networks