Production agent-infrastructure engineer who owns planning, safety, evaluation and runtime systems for a high-volume commerce agent.
Built resumable multi-loop execution, human approval gates and promoted-plan caching that cut LLM cost by 50–60%.
WebMCP, authenticated crawling and a live macOS AI product.
Resumable stateful loops plus patented execution-plan memory.
50–60% lower LLM cost and ~44% lower median latency.
Technical screen
- Design a planner–executor that can pause for approval and resume days later.
- Walk through promoted-plan invalidation, safety boundaries and rollback strategy.
Experience
3 rolesOriginal author and primary contributor to the agent runtime: 275K+ customer turns, 90+ commerce actions and 160K tool calls/month.
Built an authenticated serverless crawler and moved it to Lambda, reducing operating cost by 30%.
Built Opsgenie data-residency observability and received a pre-placement offer.
