I build the platformsthat money and datamove through.

Most teams bolt AI onto the end of their workflow: a chat window in another tab, used for whatever felt tedious that afternoon. I have spent the last two years pushing it the other way — into specification, implementation and review, where the constraints actually live and where a wrong answer is still cheap to catch.

What changes is the shape of the work rather than its volume. Writing code stops being the bottleneck, and the bottleneck becomes deciding what is worth building and judging whether what came back is correct. Neither of those is automatable, and both are the parts I would want to be doing anyway. The acceleration comes from collapsing the distance between them.

The practical lesson has been that context beats tooling. Clear specifications, typed interfaces and documented decisions matter far more than which model is answering, because a codebase legible to a new engineer is legible to a model for the same reasons. Most of the work is making the system explain itself — which is worth doing whether or not anything is reading it.