The brief

EPAM wanted to ship AI agents to enterprise clients without each client team having to re-build the orchestration layer. The platform needed to be reliable enough to sign contracts against, observable enough to satisfy security reviews, and cheap enough to be a differentiator.

What I built

• A LangGraph-based runtime that wraps every client agent as a reproducible state machine
• A Temporal workflow layer that handles long-running, multi-step agent processes with retry, idempotency, and human-in-the-loop
• An RCA subsystem that triages agent failures, classifies them into known vs. novel patterns, and routes novel ones to an on-call engineer
• An RBAC model that lets client admins control who can invoke which agents and what data they can touch

What I learned

The hardest part wasn't the agents — it was the contracts. Every client wanted different SLAs, different data residency, different observability backends. We built the platform to be the same everywhere and the contracts to be the customisation surface.

← Back home