I build LLM systems where the model proposes and deterministic code decides.
My projects are agents that work close to something expensive to get wrong: refunds, a company's data warehouse, its policy documents. The model handles language. Rules engines, gates and a human approver make the decisions, and evals check both.
I'm looking for an AI engineer role building agents, retrieval systems and the evaluation around them.
Triages support tickets against company policy, decides refund eligibility with a deterministic rules engine, drafts a reply, and holds everything for a human to approve. There is no code path from a model tool call to a refund.
The first eval run missed 5 of the 15 tickets that should have reached a human. Four fixes later it misses none. It runs live on AWS, deployed from Terraform.
FastAPI · LangGraph · Postgres + pgvector · Claude · Terraform · AWS
How an LLM support agent can propose refunds without being able to issue one. Four independent barriers, any one of which is enough on its own.
An AI data analyst that turns plain-English business questions into validated SQL, charts and cited reports, without ever giving the model a database connection.
FastAPI · SQLGlot · DuckDB + Postgres · OpenAI · Next.js
A multi-tenant RAG platform where every claim cites the exact document version and chunk it came from, and fabricated citations are stripped before the answer is returned.
FastAPI · Postgres + pgvector · hybrid search · reranking · Next.js
