Building an AI Assistant That Actually Understands Your Incidents

When production breaks, the clues are everywhere but the story is scattered. RUM sessions, APM traces, metrics, and logs allow Datadog to tell you part of the story. What’s missing is your service ownership quirks, your team's conventions, or the three-year-old workaround buried in a comment that explains why a certain endpoint behaves the way it does.

Scott Gonyea, Staff Software Engineer at Figma, will share how his team built AI-Trace, an internal debugging assistant that pairs Datadog's MCP server and Slack Incident Threads with Figma's own code repositories, architectural design, and engineering conventions to help engineers resolve user-reported issues faster. He'll cover how AI-Trace correlates RUM, APM, and log data to surface likely causes, how it grounds its answers in Figma-specific context, and the design decisions that enable it to work as a true assistant rather than just a chatbot.

Scott will walk through the architecture, access-boundary decisions, and runbook-like skill design that make AI-Trace a debugging partner instead of another dashboard. Attendees will leave with a practical pattern for combining observability data, internal engineering knowledge, and AI to reduce incident resolution time.