I build the systems that sit behind AI products, and I have been doing it for almost five years. Most of that work is real-time, the kind where latency is measured in milliseconds and things break in front of real users.

Building the model is hard, but so is everything around it, and that part gets far less credit than it deserves. Making it run fast, moving audio and video in real time, and keeping the whole thing standing when real people show up is its own kind of problem. That is the work I have spent my time on, the part nobody really notices until it stops working.

Over time I have come to think AI is becoming a systems problem as much as a modeling one. Better models matter, but what you build around them is starting to matter just as much, whether it is fast, whether it stays up, and whether it works in real time.

The more I sit with it, the more I think even the modeling problems are software engineering problems in the end. I keep hearing the people who actually build the models say the same thing, that after all the training and the research, it still comes back to engineering, to getting the system right around the model. The deeper you go, the harder it is to tell the two apart.

So that's the direction I'm going. Right now the work is real-time AI systems. Over time I want to go deeper, into machine learning systems, world models, and robotics, where the line between systems engineering and machine learning starts to disappear.