clearRx

Drug–drug interactions are exactly the kind of thing a computer should catch, but a lookup is only as good as its retrieval and its honesty about uncertainty. With Ashwin at HackGT, I rebuilt the assistant from a fragile FAISS name-lookup prototype into a real, evaluated RAG system — dense plus hybrid retrieval, cross-encoder reranking, and answers streamed with their grounding — so the interaction risk it reports is actually defensible. I focused on the retrieval and evaluation side.

Stack

• TypeScript
• Python
• BGE embeddings
• cross-encoder reranker
• SSE