Justin Karr

I spent five years in the Marine Corps and sixteen at NGA doing geospatial intelligence, through three combat deployments. For most of that I was the person using the tools, not building them — until I got tired of doing by hand what a script could do and started automating my own job. That turned into thousands of analyst hours handed back to the people who needed them.

I build AI now, at a different scale but for the same reason. Production systems with real users and real failure modes: a multi-agent platform running 30+ pipelines a day, evaluation harnesses, and the unglamorous plumbing that keeps the whole thing honest. Having depended on tools that looked impressive and fell over when it counted shapes how I build ones that don't.

I hold my own work to that standard. My independent trading research pre-registers every hypothesis and throws out the ones that don't survive honest assumptions — I've killed far more ideas than I've kept, and that's the point. A system that tells you what you want to hear is worse than no system at all.

I'm at my best when the problem is ambiguous, the stakes are real, and the thing actually has to ship.