Thirty thousand edge cases

Seventeen years. Thirty thousand customers. Every one of them came in with a different problem, a different machine, a different story about what happened right before it stopped working.

There’s no curriculum for that. You learn by doing the work, every day, with real consequences. When you get it wrong, the customer loses their files, their payroll, their manuscript. You don’t get to ship a bugfix. You fix it now, in front of them, or you don’t fix it.

That experience shapes how I think about AI systems in ways I couldn’t have anticipated.

Most AI systems are built by people who have never been responsible for something that had to work for a real person in a real moment of need. They’re optimized for demo performance, not operational reliability. They’re tested on happy paths, not on the messy, ambiguous, incomplete inputs that real people actually provide.

I build differently because I’ve seen what happens when systems fail people who can’t afford for them to fail.