Ask My Portfolio
Anyone can build a RAG prototype. The hireable signal is doing it rigorously — citations, honest refusals, and an eval harness — so it reads as a product, not a toy.
The problem
Recruiters skim. I wanted them to be able to ask my portfolio anything and get accurate, sourced answers — without me being there, and without an AI confidently making things up.
My approach
An authored corpus of my projects is chunked and embedded; a question is embedded and matched by cosine similarity, and only if retrieval is strong does it generate a citation-grounded answer (Claude first, Gemini as fallback). Below the confidence threshold it declines. An eval harness scores retrieval hit@k, correct-refusal, and answer relevance against a golden set, with a k-sweep experiment — so quality is measured, not assumed.
The result
Live, public, and free to run: 95% answer relevance and 100% correct-refusal on the eval set. Try asking it about any of the projects on this page.