Started with cameras, drones, and DaVinci Resolve. Now I build tools — mostly for the gap between AI demos and the real world: phones, watches, Windows, video pipelines. Built with AI; the part I care about is what survives outside the editor.
Yoru Studio — Content studio for one. $5/month VPS, AI agents over MCP, append-only write.
Akari Pulse — Reverse-engineered health data off a vivo WATCH GT. The official path won't export it, so: watch RPK → phone APK → own database → MCP.
LyricLens — Song lyrics → flashcards, synced to playback. NetEase plugin + standalone desktop app in Tauri/Rust.
is-ai-down — "Is it just me?" One page. Status + reachability probes on Cloudflare. Live.
chatgpt-mcp-connect — Getting an MCP server into ChatGPT is real but barely documented. This is the path that actually works.
Claude Desktop Kill-Switch — Caught it connecting directly 25 times in 12 seconds when the proxy dropped. Fail-closed firewall rules. The README walks through six "fixes" that look correct and aren't.
Motion System — 18 Remotion templates from 636 manually curated references. Timing measured from renders, not vibes.
Blender Story Video — Agent Skill: one idea → 4K H.265 through Blender + Resolve. Not shipped until frame count, encoding, and checksum pass.
Showcase Video — Agent Skill: project materials → video plan. Every unverified claim flagged before a word of promo is written.
- Operit — Android AI agent · 8 merged PRs: Compose timing, model integration, tool APIs, crash handling
- OpenCLI — Codex Desktop CDP adapter (#2281)
- open-kimi-ppt-skill — browser launch + download race (#5, #6)