Field notes from operating AI systems.
Not announcements, and not thought-leadership for its own sake. What we've learned drawing the line, building the guardrails, and watching systems after launch.
The writing
Short, specific, and written by the people doing the work. Each piece comes out of a real engagement or a real pattern we've seen often enough to be worth setting down. If a post doesn't teach something we actually learned the hard way, we don't publish it.
The articles.
We publish when we have something worth saying, not on a schedule.
Why AI systems fail in month three
An AI system that performs well in a demo and loses reliability in production is not suffering an AI problem. It is suffering an engineering problem with a specific, identifiable cause.
Drawing the line
The most consequential decision in an AI build is also the least visible: where the line goes between the parts that must be exact and the parts that should use judgment.
Monitoring has to watch a new thing
Traditional monitoring asks whether a system ran and whether it was fast. An AI system also has to be watched for whether it is still right.
More as we write them.
We publish when we have something worth saying, not on a schedule.