I will build an ai agent workflow with rag evaluation that does a real job

Most agent demos work once and break the first week they meet real data. I build the parts that make them hold up.

A human approval gate, so nothing irreversible happens without sign-off. A validation loop that sends work back until it converges, instead of accepting the first plausible answer. An evaluation harness, so when you change a prompt or swap a model you can tell whether it got better.

Multi-agent orchestration when one agent is not enough: a proposer works, an independent validator rejects what breaks the constraint, and it comes back corrected.

If your agent answers from your own documents, that is RAG, and the hard part is knowing whether retrieval returns the right passage. I build the RAG evaluation with it, on a held-out set you rerun after every change, so quality is a number, not a feeling.

8 years as a backend and data engineer at an enterprise data company: 439 pull requests reviewed across 19 repositories, 107 incidents resolved.

You get the workflow wired into your tools, the harness, source code and docs. Java, Python, TypeScript. GPT, Claude, Gemini.

Tell me the job and I will say honestly whether an agent is the right tool.