Sports Science × AI Product Builder × Climber
I turn real-world sports workflows into human-in-the-loop products. My starting point is usually not a technology trend, but a friction I have seen in a training venue, competition workflow, or research process.
我是一名运动人体科学研究生，也是一名场景型 builder：先和真实用户把问题说清楚，再用产品拆解、数据与 AI 协作开发，把想法变成可以被试用、被质疑、被改进的工具。
A desktop workflow for gymnastics competition video operations:
long-form video → AI candidates → human review → score-card binding → batch delivery
- Built with Electron, React, TypeScript, FastAPI, Python and ffmpeg.
- Keeps AI in the candidate-generation role and people in the final-decision role.
- Uses short-lived branches, Pull Requests, CI, structured Issues and versioned Releases.
- Current verified baseline: 257 backend tests passed / 59% coverage.
Repository · Releases · CI
- Start from the field — understand the person, physical context, risk and current workaround.
- Translate before building — turn stories into user flows, PRDs, data boundaries and acceptance criteria.
- Build a tracer bullet — test the most uncertain vertical slice before expanding the system.
- Keep evidence — Ground Truth, reproducible tests, reviewable state and badcase-driven iteration.
- Design a way back — explicit fallbacks and human confirmation for risky automation.
Product discovery · PRD & system design · Human-in-the-loop evaluation · Python · FastAPI · Electron · React · TypeScript · Codex / Claude
- M.S. candidate in Human Kinesiology at the China Institute of Sport Science.
- Former university swimming team captain.
- Climber and certified wilderness first aider.
- Interested in sports technology, embodied interaction and tools that make complex work feel simpler.
I enjoy building a polished demo, but I care more about making a real person do one less frustrating step.