Overview
Build a biomedical QA system where each generated claim remains traceable to retrieved evidence and easy to inspect during evaluation.
System design
- Sparse and dense retrieval over a 26.8 million document PubMed corpus
- BM25, Pyserini/Lucene, MedCPT, TF-IDF/MMR, FAISS, and cross-encoder reranking
- Qwen-based generation served on GPU infrastructure through vLLM and Slurm
- Human and AI feedback loops to refine reformulation, ranking, and answer generation behavior
What was built
- End-to-end pipeline for TREC BioGen Task-B submissions
- Citation-linked answer generation with PubMed evidence traces
- Interactive dashboard for comparing runs, inspecting outputs, and debugging evidence quality
Results
- Contributed to a published TREC 2025 system paper
- Produced reusable diagnostics for answer accuracy, retrieval quality, evidence grounding, and citation reliability
- Made system behavior easier to inspect for both manual review and shared-task iteration