I build control systems, and lately I build software the way I was trained to build control systems.
Twenty years in building automation and integrated building systems: the controls, physical security, fire, and networks that run a building, and the commissioning discipline that proves they do what the drawings say. Before that, the Navy's nuclear program, where you assume nothing works until you have watched it work.
Now I build software with AI coding agents every day, and the same discipline turns out to be the missing piece. An agent will report a task complete when it is not. A control nobody has watched deny anything is an assumption. A plan regenerated from the code launders drift into the record. None of that is new. It is the failure mode every commissioning engineer already knows, arriving in an industry that has not learned to distrust a status point yet.
The projects below are what that conviction produced.
A governance methodology for building software with AI agents. A human ratifies intent, work is dispatched one bounded unit at a time with an explicit capability grant, a separate review checks the result against the record, and every enforcement layer is birth-tested before it is trusted. Public, with the full self-hosting pilot log, including the measurements that went the wrong way.
An open-source, deterministic parser for construction drawings and specifications. It reads a drawing set and a spec book the same way every time, with no model in the loop, because a document-control pipeline cannot rest on a component that guesses.
A pure-Rust audio analysis crate, published on crates.io. Signal analysis with no runtime dependencies and no surprises.