Thu, Dec 1011:00 AM PST, 30 min
SRE agents you can trust with on-call and incidents, requires beating even the frontier models on accuracy, latency, and cost, all at once.
Getting there requires a research lab and it led to two breakthroughs. First is SRE-1, our first model that dominates the Pareto frontier of accuracy against cost on root cause analysis of production issues, beating frontier models.
Second, ResolveAI is 2x as accurate as coding agents at 2x the speed, so it can now sit on the critical path, closing issues on its own and bringing in engineers only when needed.
Join Dhruv Mahajan and Sean Bell from ResolveAI Labs for a 30-minute tech talk on how we built it and what’s next.
Dhruv Mahajan
Chief AI Scientist, ResolveAI Labs
Sean Bell
Member of Technical Staff, ResolveAI Labs
Production is dynamic and issues are ephemeral. Training models requires building the right environments to learn from
Models learn to game whatever grades them. Reinforcement learning requires continuous rethinking how SRE-1 is graded and corrected
Most incidents never get a written answer to check against. How we approached verifying and measuring SRE-1 against frontier models
ResolveAI has to work under time budget and quality constraints. How we harness intelligence to continuously stay on the frontier