AI and Risk Mitigation: Call for Proposals

Lumian runs LLM agents in production for Amazon sellers - real customer data, multi-tenant isolation, heavy tool-calling. This talk covers the risk-mitigation side of that: designing tools agents cannot misuse, building regression harnesses that catch model drift before customers do, and the failure modes that only show up once real money and real data are on the line.

Practical, code-level patterns for anyone shipping agentic systems who cares about safety over demos - drawn from building Lumian’s agent platform end to end as a founding engineer.