Writing

Notes from building AI systems that have to survive real users — mostly about the engineering around the model rather than the model itself.

What it takes to trust a model with someone's books

Getting document extraction to a number you can defend — and building the layer that catches it when it is wrong anyway.