Flagship · RAG + MCP
Ask about Kumail's work
Answers come only from his resume and project READMEs, with citations. Built with hybrid retrieval (BM25 + embeddings) and an MCP server.
86.4% BM25 hit@1 · 22-question eval100% Embeddings hit@135 Tests · CI green
Example questions
Evidence · Ask Kumail
Measured, not claimed
Hit@1 on a 22-question evaluation: did the first retrieved result contain the answer? BM25 from scratch gets most of the way there — embeddings close the gap, and fusing the two is what ships.
86.4%
BM25
90.9%
Hybrid
100%
Embeddings
| Method | Hit@1 | Eval size | Note |
|---|---|---|---|
| BM25 (written from scratch) | 86.4% | 22 questions | Keyword ranking, no libraries |
| Hybrid | 90.9% | 22 questions | Reciprocal rank fusion |
| Embeddings | 100% | 22 questions | Semantic ranking |
| Reciprocal rank fusion | Ships in the product | — | Merges both rankings for the final answer |
Use it from Claude or any MCP client
Add this as a remote MCP server. Tools: search_docs, ask_kumail, list_projects.
/api/mcp