About

I've spent about ten years building backend systems: Go microservices, GraphQL APIs, event pipelines on Kafka, and infrastructure on AWS and GCP. Most of that work is about getting correct answers out of systems that run continuously and fail in partial, inconvenient ways.

I'm now moving that into AI/ML and robotics. Those habits still apply there: fixed timesteps, latency budgets, tests that prove one component matches another, and not trusting a good number until the evaluation behind it holds up.

What I'm working on

• Legged locomotion in simulation. The Go2 demo runs a policy I trained with MuJoCo Playground, then exported to a small network that runs in the browser.
• Evaluation you can trust. nautilus-alpha is built on the idea that the protocol matters more than the model.
• Vision to prediction. ufc-fight-vision turns fight footage into style features.
• Agents with guardrails. direx lets AI agents draft marketing work and requires a person to approve everything before it goes out.

I don't publish a resume or contact details here. My work is on GitHub.