⌘projects04

a production rag system over 111 indian startups with a hand-rolled retrieval pipeline — pgvector cosine search + postgres full-text search fused via reciprocal rank fusion, reranked by a bge cross-encoder (no langchain). custom llm-as-judge eval harness (reranking lifted mrr 0.688 → 0.750; 0.94 faithfulness on the golden set), streaming sse chat with grounded [source n] citations, langfuse tracing, and a retrieval lab comparing vector vs hybrid vs hybrid+rerank. live on gcp cloud run + vercel + supabase.

a provider-agnostic async python client that normalizes messages, streaming, token usage, and errors across openai, anthropic, and gemini behind one interface. ships as a library, cli, interactive repl, and fastapi service — with tenacity retries, pydantic v2 models, mypy-strict typing, a concurrent compare() across providers, and 40 tests.

built a citation-style rag chat app where users upload pdfs and ask grounded questions. pipelines gemini-embedding-001 into pinecone (namespace per file), streams answers via gemini-2.5-flash, and stores files on vercel blob with neon postgres + drizzle orm.

created a chatbot with gemini api for conversations, tested over 100 sessions to refine response accuracy. added clerk login and mongodb for chat storage. used tailwindcss for a mobile-friendly ui.