I work in the layer between a model's API and the person waiting on the other end.

I've spent about three years building Node and TypeScript backends. Somewhere in there I realized the prompt is rarely the hard part of an AI product. The hard part is everything around it. Keeping background jobs from taking each other down when one of them fails. Holding onto a user's history so it still makes sense to the model months later. Getting responses to stream so the app feels alive. Making the auth hold up when someone actually pokes at it.

I care most about the parts of a system you never notice when they're working. I keep orchestration, tool calls, and storage cleanly separated, and I do it for a practical reason: it lets me think through what happens when something fails before it actually does. I try to build things that stay calm when they break. And I try to leave enough of a trail that the next person can follow why I did what I did.

I don't do model training or MLOps. I work on the backend that sits between the model and the product, and I think that's where a lot of AI apps quietly succeed or fail.

AI has made me a lot faster, but it hasn't done the thinking for me, and I don't expect it to.