AI AgentsPrivate
Production AI Backend
Client work — empty repo to shipped
The first time I carried something this involved all the way to production for a real client. I owned it end to end — schema, auth, retrieval, orchestration, evaluation, deploy.
The interesting constraint was cost. Running every request through a frontier model would have been simple and unaffordable, so classification routes to a cheap model and only generation escalates. The eval harness exists so I could prove that tradeoff didn't quietly degrade quality.
What it does
- FastAPI service with JWT auth and row-level security over Supabase/Postgres
- Hybrid RAG retrieval over pgvector
- Three-tier persistent-memory layer, so conversations remember you across sessions
- Hand-rolled agent orchestration loop instead of reaching for a framework
- Cost-aware model routing: cheap model for classification, stronger one for generation
- Eval harness so quality gets measured, not eyeballed