RL Environment Engineer
Bespoke Labs
- Developed AI agent evaluation tasks for Nebula Aurora, a Kubernetes-based benchmark that tests frontier AI models on real DevOps and SRE incidents
- Wrote setup scripts that inject controlled failures into K3s clusters, simulating incidents across CI/CD pipelines, service mesh configurations, and observability stacks
- Built Python graders with partial scoring that check agent responses against actual system state, plus ground-truth solution scripts; every task had to be verifiable from observable system state alone
- Annotated coding-agent trajectories to produce training data for agents' research and planning ability