Research
Reliable agents, grounded environments, and AI systems
My long-term direction is to understand how AI agents can be trained and evaluated in environments where reliability is measured by system outcomes.
Research focus
I am most interested in agent behavior that emerges across a sequence of observations and actions: diagnosis, tool selection, intervention, verification, and recovery.
Mechanically verified agent environments
Reinforcement learning for infrastructure repair
Agent evaluation and failure analysis
Post-training methods such as SFT, DPO, GRPO, and RLHF-style environment feedback
AI systems and infrastructure for training/evaluation
Current project
CrashDiag
CrashDiag is a mechanically verified infrastructure-repair environment for training and evaluating agents. A model receives an incident observation, emits one bounded JSON action, and the environment checks whether the sandbox state was actually fixed.
Mechanically verifiedInfrastructure repairGRPO
Future directions
Benchmarks where reward comes from executable system state, not prose grading.
Training signals from structured sandboxes, health checks, and operational feedback.
Failure taxonomies for diagnosis, remediation, recovery, and verifier design.
Reusable infrastructure for reproducible agent training and evaluation at research scale.
Reading and research notes
A curated notes section will be added as the reading list and project logs become more structured. For now, research diary posts live on the blog.
View blog