Het Patel

Generalist leaning full-stack plus applied AI. Python for ML and eval work, JavaScript/TypeScript for the web side. Comfortable across the RAG stack from retrieval to evaluation harnesses.

RAG evaluation that actually holds up: claim-level faithfulness with NLI entailment, where token-overlap metrics break, and how to separate faithfulness from relevance.

A RAG eval harness, currently stuck on the claim-extraction step where over-splitting and under-splitting both quietly wreck the faithfulness score.

Talking shop about RAG eval, retrieval pipelines, and LLM failure modes. Also open to ML and full-stack roles in Germany. Say hi if you spotted a hole in my faithfulness post.