Research
My graduate work focused on evidence-grounded biomedical question answering: designing systems that retrieve well, answer carefully, cite correctly, and reveal where they fail.
Evidence-Grounded Biomedical Question Answering: Design, Analysis, and Evaluation
Published through Concordia University's Spectrum research repository. The work joins biomedical retrieval, reranking, language-model generation, citation validation, and diagnostic evaluation into one system view.
BioEvidence
The research pipeline became a working application with independently inspectable stages, biomedical API integrations, Supabase authentication and data, and a Next.js interface.
A complete evidence path
Find the right documents, chunks, and signals before generation starts.
Shape context so the model stays tied to evidence instead of drifting.
Use LLMs where they help, but keep outputs readable, constrained, and useful.
Measure failure modes, citation quality, and where trust breaks down.
Evidence-grounded QA
Retrieve and rerank evidence before generation, then verify whether the final answer is supported by the cited sources.
Stage-aware evaluation
Separate retrieval, answer quality, grounding, and citation behavior so failures can be traced to the right part of the system.
NLP under imbalance
Compare BERT-family, CRF, hierarchical, and PPO approaches using per-label behavior and detailed error analysis.
Repeatable inference
Run quantized language-model experiments through vLLM and Slurm on V100/A100 GPUs at evaluation scale.