I'm a CS and Stats student at UC Berkeley. Most of what I build is about one question: how do you know an AI did the right thing? So my projects tend to come with a test harness, a replay, or a benchmark attached.
🏆 1st place, AIE x MLH Hackathon @ Code Summit (Nov 2026) for Muscle Memory
Website · X / Twitter · Berkeley, CA
| Project | What it does |
|---|---|
| Accord · live map | A map of 24,063 perfumes placed by their ingredients. Find what smells like what you love, and a cheaper bottle with the same notes. |
| Muscle Memory · 🏆 1st place | An editor that cleans up AI-written code. It reasons slowly at first, then turns calls it keeps getting right into instant Semgrep rules. Every cut is checked against hidden tests. |
| inference-bench | When does speculative decoding actually speed up vLLM? N-gram vs draft-model speculation on code edits vs open-ended text, across batch sizes. |
| Airlock · replay | Forks microVM sandboxes to find the exact conditions where an AI tool leaks secrets, then tests fixes. |
| Clean Room | An AI that rebuilds software by watching it run, then proves the copy behaves the same on tests it never saw. |
| Probation | A strict backtester a trading agent has to pass before it touches real money. No lookahead, real fees, realistic fills. |
- Languages: Python · TypeScript · C++ · SQL
- AI: vLLM · Gemini · Claude · agent tooling and evals
- Web: React · Next.js · Expo · Convex · Vercel
Currently building at hackathons around SF most weekends. Always up for talking about evals, agents, or weird data sets.

