Expanding METR's open-source Vivaria platform for running & evaluating sandboxed LLM agents, adding support for lightweight branches of run trajectories, GPU allocation, multiple LLM providers, etc.

- Productionized an initially research-grade VSCode extension codebase, integrating Microsoft's experimentation framework, factoring out reusable editor-agnostic core library now powering the IntelliJ, VS, and Neovim integrations.- Added fast & extensible stream filtering and completion-stopping logic to our Go server, then ran full stack experiments verifying utilization and acceptance rate improvements.- Optimized model utilization throughout GitHub, MS and OpenAI parts of the serving stack.- Onboarded engineers as team grew 10x; improved dev-ex with debug & visualization tools.

- Tech lead for small team developing Cider, Google's internal-only web-based IDE.- Added multiple version control systems, debugging, custom keyboard shortcuts, etc.- Scaled the codebase to 300k LoC, migrating to TypeScript, adding Angular components &state-management, and creating extensibility APIs for contributors from other teams.- Grew Cider into the most widely used IDE for Google's multi-billion LoC monorepo.

...

...