I turn ambiguous ideas into AI products people can use, inspect, and trust.

Graded Chinese listening from real news, with synced pinyin and translation. Web, iOS, and podcasts.

• I shipped the learning loop, so stories, audio, and vocabulary all match each learner's HSK level.
• Episode production runs as an agent workflow I codified: every episode clears a vocabulary check against that level, editorial and LLM review, and a human listening pass before it publishes.

An AI coach for amateur athletes. It reads their training data and writes plans they can understand and change. Chinese beta.

• I own the agent: tools over activity data, a coach that proposes and revises plans, and the evaluation harness that grades those suggestions.
• I built most of the product it sits in — activity import, planning, teams, the workspace — and beta-tested it with 30 active runners across two university running groups.

Let people and external agents create, iterate, and deploy applications through an interface they can trust.

• I took Atoms CLI from 0→1, giving coding agents a direct path to create, inspect, validate, and deploy applications.
• I defined a full-stack data workspace so non-technical builders could inspect and update app data without SQL.
• The platform was charging users to repair its own agent's mistakes — 34% of bug-fix requests — so I made the case for tiered pricing.

A cross-platform timer that makes breaks harder to skip while keeping its rules transparent.

• I designed it for Windows, macOS, and Linux around transparent rules and explicit user control.
• Seven years on, it's multilingual, still maintained, and kept alive by contributors I've never met in person.

An AI-assisted paper-cutting tool that helps creators work with cultural symbolism, not shape alone.

• I designed a system that connects visual form with cultural meaning, then evaluated it in a 16-person study.
• We published it as a full paper at ACM CHI 2025, where I'm co-first author.

A computer-vision defect-detection system that learned from normal production samples and ran in real time on low-cost edge hardware.

• I co-led it, pairing normal-only anomaly detection with a dashboard I shaped by watching the factory floor.
• I optimized the vision pipeline for real-time inference on roughly $40 of hardware and deployed it across 20 loom stations, where operators clear defects and managers watch production status.