Most of what I build starts as a problem in my own life. The fix is usually more interesting than the problem, and that's how the list got this long.
🐋 MemWhale. My coding agents kept forgetting what they'd already learned, so I built a local-first terminal memory system that records everything you put in and serves it back over MCP.
🎵 Cadence. AI writing has a tell, so I built a tool that catches it and rewrites the draft to sound like a person. Runs in Claude Code, Codex, Gemini, DeepSeek, the browser, and the command line. It also trains its own humanizer: LoRA-Cadence, a rank-16 QLoRA on Qwen2.5-3B-Instruct, graded by Cadence's own slop detector.
🎬 Real. Veo 3 kept ignoring half of what I asked for, so I built a tool that assembles video prompts from ones that worked before and holds the output to what I actually specified.
📡 WireDAQ. You shouldn't need the hardware to start, so I built a data-acquisition simulator: test the packet protocols and ground-station tooling first, then swap in real hardware when it's ready. Python and C share the codecs.
🎻 BowTrack. To fix my own cello bowing, I built a computer-vision tool that reads bowing motion, posture, and early technique signals with OpenCV and MediaPipe.
Also ongoing
🌊 Localization, estimation, and mission planning for Barracuda 2.0, an autonomous underwater vehicle, with my teammates at USC AUV.
🚀 Embedded data acquisition, sensor drivers, and high-speed telemetry for rocket testing with my teammates at USC Rocket Propulsion Laboratory.
🤝 I contribute to the tools I use. I'm a consistent contributor to open-source Rust agent harnesses like rho and CodeWhale, and I've merged security fixes into Jan and LocalAI.
📚 More: