Peptide AI RAG and Agent Platform

The problem

Users ask health questions that have to be answered from their own data, meaning their wearable signals, their logs, and their history, with no leakage between users, and it has to stream fast enough to feel like a conversation. At around 9 seconds to an answer, people stopped engaging with it.

What I built

A TypeScript retrieval stack on Supabase Edge Functions: hybrid dense + BM25 retrieval, Voyage cross-encoder reranking, streamed Claude responses, and a pgvector semantic cache scoped per user with dedicated safety guards.

Latency was the product problem

At around 9 seconds to a first useful answer, users stopped asking follow up questions. The thing that fixed it was not the choice of model. It was noticing how many questions were semantically similar to ones already asked. A pgvector cache keyed on embedding proximity absorbs those, and a cache hit returns in under 3 seconds at no extra inference cost.

Caching health questions safely

A simple shared cache would leak one user's data into another user's answer. Entries are scoped per user and gated by dedicated safety guards, and a Deno regression suite runs on every deploy to assert the negative controls, which are the questions that must never return a cache hit.

Separating computation from narration

The Coach system computes statistics deterministically in code, then asks the model only to narrate what was computed, citing evidence tiers. Numbers stop being hallucination surface area.

Grading the trajectory

Scoring only the final output hides the failures that matter. The evaluation covers claim level grounding, whether the tool trace was correct, and whether the agent abstained on negative controls. That means an agent that arrives at the right answer through the wrong tools still fails the suite.