AI product and agent-workflow builder
AI 产品与智能体工作流开发者
I build AI products and developer tools with traceable evidence and human judgment at their core.
我把研究训练与教学经验带进产品设计，让复杂工作流变得可用、可核验。
Website · Projects · Research · Screen · Role Fit
My work connects product, research, and education: from AI assistants and coding-agent tools to local-first research apps and creative workflows. I design the user journey, build the system, and carry it through verification and release.
My research focuses on Cantonese speech, tone, and emotional prosody. Experimental design, acoustic analysis, and teaching shape how I understand users and evaluate what a product actually does.
| Project | What it helps people do | Explore |
|---|---|---|
| Storyspan 源映 | Develop a director's creative intent into a screenplay, with traceable source material, revision approval, and Fountain export. Local-first; in development. | Project |
| CineTrace | Connect film shots, speech, and keyframes to evidence-linked analysis and HTML reports. Automated visual observations remain drafts for review. | Project |
| Role Fit | Compare a job description with portfolio evidence, explore supported matches and gaps, and ask follow-up questions. Live in three languages. | Try it |
| Portfolio assistant | Explore my projects and research through source-linked answers and interactive cards on my multilingual Astro website. | Website |
| Project | What it does |
|---|---|
| Evidoc | Checks documentation and coding-agent instructions against repository evidence. Available as @evidoc/evidoc, with a CLI, MCP server, local web UI, and GitHub Action. |
| D-academic-agent | Local-first workspace for checking academic claims and citations against paper evidence. |
| D-viva-assistant-agent | Thesis viva preparation with evidence-linked practice, answer feedback, and revision tasks. |
| RelayBar | macOS menu-bar utility for monitoring Codex and Antigravity account pools and quota. |
| Cantonese–Mandarin Cross-Reference | Pronunciation lookup with Jyutping, Pinyin, audio, and mobile support. |
| Dong-skills | Reusable skills for scoped agent collaboration, review, and handoff. |
Use OpenCode, Grok, or Antigravity from Claude Code and Codex for scoped tasks, code review, troubleshooting, and session handoff.
| Target CLI | From Claude Code | From Codex |
|---|---|---|
| OpenCode | opencode-plugin-cc | opencode-plugin-codex |
| Grok | grok-plugin-cc | grok-plugin-codex |
Antigravity (agy) |
agy-plugin-cc | agy-plugin-codex |
The Claude Code family traces back to OpenAI's Apache-2.0 codex-plugin-cc: the OpenCode port adapts its command surface, the Grok port applies the same approach, and the Antigravity port builds on the OpenCode port. The Codex family uses MCP servers and collaboration skills. The Antigravity plugins isolate review work from the source repository.
- Speech and language: Cantonese vocal emotion, lexical tone, and second-language speech. My work includes a first-author article in Journal of Speech, Language, and Hearing Research and a paper accepted by Language and Speech. Publications
- Film and criticism: short films made with generative models and long-form close readings of AI-native drama. Watch & read
AI & developer tools: TypeScript · Node.js · Python · MCP · AI SDK · coding-agent CLIs
Web & desktop: Astro · React · Next.js · Electron · SQLite · Swift · macOS
Film analysis: FFmpeg · PySceneDetect · faster-whisper
Delivery & verification: GitHub Actions · Vitest · Playwright · Vercel
Evidence first. Human judgment stays in the loop. Build something useful.