Before I treated code as a craft, I spent years designing and running automated trading systems in production — real capital, real latency, the daily discipline of being wrong in front of a P&L. It teaches the one thing that transfers everywhere: a system that's clever in a notebook and a system that survives contact with reality are different animals.

Now I build smaller, more stubborn things — tools that preserve, decode, and keep working when the comfortable infrastructure is gone. One quiet thread runs through all of them: give people back what was faked, locked, abandoned, or lost.

• FLAC Detective — catches MP3s smuggled into lossless files. The interesting part isn't the spectral rules; it's the ML write-up: four approaches that failed, a self-caught false discovery (a too-good "AUC 0.99" that cross-validation killed), and a "fundamental limit" that turned out to be me listening in mono. I'd rather document the wrong turns than pretend the road was straight.
• Prometheus-Station — a solar-powered Raspberry Pi serving offline Wikipedia over a long-range LoRa mesh. For when the grid and the network aren't there.
• DS2-Anywhere — decoding Olympus DSS/DS2 dictation files with no Windows, no GUI, no paid software — a format that stayed locked for ten years.
• HAP-Revival — keeping Sony's HAP-Z1ES / HAP-S1 audiophile players alive after Sony walked away in 2021.
• Creabook — a book-generation engine, for when the pen tires before the idea does.

Production-first, honest about uncertainty, allergic to cleverness that doesn't survive testing. I read FIX logs and Pascal with about equal pleasure.