AI-powered conversational medical assistant with RAG over MedLinePlus & MedQuAD datasets, Neo4j knowledge graph, Qdrant vector database, LangGraph agent orchestration, and voice I/O (Whisper STT + Edge TTS).

Fine-tuned PubMedBERT bi-encoder on PubMedQA achieving NDCG@10=0.974. BM25 from scratch in pure NumPy. Self-improving multi-agent critic loop. Model live on HuggingFace Hub.

Local multi-agent pipeline (LLaMa-3.1 via MLX + PubMedBERT RAG over Chroma) that fact-checks drug questions against authoritative medical sources and returns a cited answer with per-claim verdicts. Ships with a FastAPI + single-page web UI and a Typer CLI

Fine-tuned PubMedBERT for medical abstract sentence classification. 89% accuracy on PubMed 200k RCT

Transformer-based topic modeling (BERTopic) of six spine journals, 1978–2026 — 40,401 abstracts, 37 topics, with temporal trends and VOSviewer comparison.

Agentic Corrective RAG over 484 ClinicalTrials.gov oncology protocols. LangGraph CRAG with LLM-as-judge grading + cross-encoder reranker for two-stage retrieval. Three independent refusal gates. PubMedBERT, FAISS, Groq Llama 3.3 70B. Live demo on Streamlit.

Biotech venture valuation workbench: reflexivity-adjusted PoS, phase-stratified ML prior trained on real HINT outcomes (externally validated against CT Open uncontaminated), phase-gated rNPV with Monte Carlo, 8-pillar diligence scorecard, cited deal comparables. Modular registry architecture with three runtime agents.