I study Data Science and Cognitive Science at UC Berkeley. I build tools and guardrails for AI agents — what they cost, where they break, and what stops them from doing damage.
Outside of that, I build tooling around my own training data, because I ride a lot and Garmin gets more wrong than you'd think.
🛡️ Agent guardrails • 🤖 AI agent tooling • 🚴 Endurance data • 🔓 Open source
fly-brain — a fruit fly's whole central nervous system (166,700 neurons, male CNS v1.0 connectome) run as a spiking network in PyTorch, learning to ride a road bike in side gusts. The simulator reproduces Shiu et al. (Nature 2024) spike for spike against their Brian2 code, then batches many flies on one GPU.
token-counter — audits monthly AI coding-agent spend across Claude Code, Codex, Cursor, and Gemini, and flags every run that would have been fine on a cheaper model.
fitpet — a Claude Code status-line companion that grows from real Garmin training load.
vibeapply — an AI-native job-application copilot: a browser extension that rides along on job boards, backed by a structured candidate profile and evidence bank. Scaffolded with agents, then audited — all nine findings fixed before any feature work.
cal-dining-scanner — watches UC Berkeley dining menus for the foods you care about and emails you when they appear. One command to set up, runs free on GitHub Actions. A friend spun it off into Cal-Bite, a full-stack version.
magene-garmin-fix — Garmin inflates climb 10–70× for third-party bike computers by throwing away barometric altitude. This fixes it.
analyze-cycling-hr-thresholds — a Codex skill that reconciles Garmin field data with raw CPET gas-exchange data into heart-rate thresholds, Garmin-ready zones, and labeled FTP proxies. It analyzes each source on its own and refuses the tempting shortcuts — no converting heart rate straight to watts, no treating a lab report's conclusion as raw evidence.
bay-area-venture-map — interactive map of the Bay Area venture cluster; filter by AUM, stage, sector, and check size.
qosmic-audit-harness — contracts-first agent harness: every claim in a generated audit must cite a real crawled page, enforced by a validator that the writer and the evaluator share.
Taxuspt/garmin_mcp — 1.3k ★ · #2 contributor · 21 PRs merged
The most-used MCP server for Garmin Connect. Behind the maintainer, I'm its largest contributor: 21 merged PRs (+3.3k lines) spanning auth, FIT decoding, training analytics, workout scheduling and write safety. The pattern is the same throughout: find where an agent silently gets a plausible but wrong answer, prove it with a real payload, and ship the fix with a regression test.
| #437 | Analytics assumed 1 record = 1 second. With smart recording (a record every 1–8 s), a 20-min best of 300 W came out as 167 W, and climb VAM was off by 3×. Power curves, climbs and HR drift now use real timestamps. |
| #423 | Told the agent a retry was safe when it wasn't. A 5xx that arrives after a write was sent may already be committed. It's now marked write_committed: "unknown", so an agent checks before it double-logs. |
| #419 | Logged into the wrong account. With GARMINTOKENS set, password login silently reused another profile's tokens and wrote them into the new directory. Credentials now always win. |
| #418 | Di2 shifts decoded with the bytes swapped, so a 50×21 read as "50/2t". The tests passed only because the mocks used the same wrong layout. Rebuilt against the FIT profile, and added L/R power balance. |
| #254 | One stalled Garmin call froze the whole server for about 4 minutes. Every proxied call is now bounded on a worker thread, with a timeout an agent can retry. |
| #426 | Memory leak in a login failure. Re-raising a stored exception grew its traceback without bound: 3,000 frames after 1,000 tool calls. |
All 15 others
- Courses #435: 5 of 8 sport types were mapped wrong (a gravel course saved as track running), GPX output was unescaped, and deleting now needs confirmation.
- Analysis #436: bike power had leaked into running power curves, and HR drift is now labelled by its sign.
- Workouts #433 · #434: one failed
schedule_weekentry no longer hides the ones that succeeded; strict date validation and full library paging. - Health, training and weight #429 · #427 · #430: readiness fields, lactate-threshold history, weigh-in timestamps, and deletes that silently did nothing.
- Gear, devices, PRs and nutrition #250 · #431 · #424 · #432 · #428: gear notes from the v2 API, real activity typeIds, the actual solar payload, when each personal record was set, and rename-safe food edits.
- Null-safety #253: Garmin sends an explicit
nullfor empty sections, so.get(k, {})returnsNone. Fixed five crashes of this kind, each with a regression test. - Auth and test infrastructure #425 · #416: token verification for profiles without a full name, and live e2e tests made opt-in.
NousResearch/hermes-agent — 251k ★
- #64036 — reasoning-visibility
leak on Gemini/Vertex: with
show_reasoning: false, users still got dozens of raw thought-summary blocks. Two causes — effort and visibility were coupled in the request builder, and Vertex returns summaries untagged unless you setthought_tag_marker, so nothing downstream could key on them. Splits the two knobs to match Google's own API. - #68524 —
web_extracthad one backend; expired credits or a rate limit meant total failure plus a gateway restart. Adds a configurable fallback chain.
SpencerLiang0114/RouteCraft — route planner for running, hiking, and cycling
- #15 ✅ merged — session auth, saved routes owned by the user who created them, and share links that expire.
- #14 ✅ merged — imperial conversions next to every metric route value.
Berkeley, CA — usually on a bike when I'm not here.