I build LLM agents and the infrastructure that makes them safe to run. My MCP Gateway on Kubernetes brings three upstream MCP servers behind one authenticated endpoint, with role-based permissions, security controls, and audit logging.
For the Aster & Row Support Agent, I built hybrid retrieval and a tool-calling customer support agent. On a labelled benchmark, retrieval improvements raised Recall@3 from 83% to 100% and MRR from 0.75 to 0.90.
I also built an evaluation harness for a policy-bound support agent, combining deterministic checks with a calibrated LLM judge. Its latest run passes 24/24 cases, and I’ve shipped tested and documented projects on GitHub.
