Software Engineer Intern
Redmond, WA
—
- Root-caused and fixed 2 production segfaults in the Cowork agent loop, eliminating dozens of daily container crashes and restoring runtime stability.
- Proposed and shipped core evaluation-pipeline improvements — a standardized judge, unified case format, and dashboard — adopted as team standards.
- Solely owned Cowork's gRPC/MCP integration and 215-task paired evaluation of a new document backend behind a feature flag; repaired ~10 cases and cut false-negative failures from ~80% to 1–2% of failing cases.
- Built and iterated a React/Vite evaluation dashboard used by ~20 engineers; presented 189-vs-174 results twice to ~12 Distinguished/Principal Engineers, driving backend selection for a new product feature.