building ai agents that actually ship, not just demo well ⚡
i work on the unglamorous 90% nobody posts about: grounding, guardrails, evals, the reason your agent doesn't confidently do the wrong thing in prod. currently at servicenow, owning a langgraph multi-agent system that closes thousands of support cases a month with zero humans in the loop.
things i actually believe
the eval harness matters more than the prompt an agent that knows when it doesn't know beats one that demos clean most performance wins are hiding in code nobody bothered to profile
stack i live in
langgraph pytorch kafka fastapi kubernetes neo4j onnx
status open to ai / agent-eng roles on teams that move fast. startups especially. if you're building something where the agents have to be right, we should talk.
hit me up · ajaysai.6601@gmail.com