Joey Musselman
Software engineer building AI that's actually useful, and proving it works.
I'm in Charlotte, with five years of shipping product software, now focused on AI. I don't believe a number until it's been checked against something the model never saw.
Measuring whether AI actually works
Pass lines written before the test, misses published. My podcast checks itself, and I test sports video models blind on matches they've never seen.
Applied ML on real video and audio
Messy input, not benchmark datasets: broadcast sports footage, and a custom wake word scored on the device's own engine before it shipped.
AI people actually want around
Built for people I know: Clown Cast, a voice assistant for my house, and an archive for my old college circus.
How I work
- I learn in public, after the fact: each thing I learn gets a podcast series, written only after the work happened, with corrections on air.
- AI agents write most of my code. I write the rules they work under, in code rather than prompts, and I make the calls.
- The builder can't grade its own work. Tests come from the spec, not the code.
- I credit people by name, and nobody from my house or family appears in public work without saying yes.
At work
Software engineer at Accenture. I own evaluation for an agent system and built its back-testing harness. Smaller things I've built are on the projects page.