Notes on AI document processing.
What actually works on real documents, what quietly fails, and what it costs to run — written by the engineers building the systems, not by a marketing team.
Articles
The big-bang rewrite keeps failing. Move the system unit by unit instead.
Why full legacy rewrites go dark and come back unprovable, and what a mapped, test-backed, AI-assisted migration looks like when the system stays usable.
How to scope an AI proof-of-concept so it ends in a decision
A proof-of-concept should end in a decision your budget holder can defend, and that depends almost entirely on how you scope it before any code is written.
What a 98% accuracy number actually means
A headline accuracy percentage tells a buyer almost nothing on its own, and here is what to ask a document-extraction vendor before you take one seriously.