Training grounds for computer-use agents
From August 2025 to February 2026 I was a software engineer at Matrices, building reinforcement learning environments for the top AI labs in the world.
The environments I built train multimodal, LLM-based agents on realistic computer-use tasks: replicas of the software people use every day, where agents practice real work and labs measure how well they do it. That work was covered by The New York Times.
- Contribution
- Software Engineer
- Timeline
- Aug 2025 - Feb 2026
- Industry
- AI research infrastructure
- Backed by
- Index Ventures
- Location
- San Francisco

Environment design
I built replica apps and websites with the state, quirks and edge cases an agent meets in the real world.
Tasks and graders
I wrote task suites and automatic checks that turn an agent's run into a reliable training signal.
Lab delivery
I packaged environments so research teams could run them at scale inside their own training stacks.




