Insights
Practical notes from building systems that have to hold up.
Writing about evaluation, security, delivery, process, and the operating decisions that turn an AI capability into a dependable service.
Why a named person's login should never authenticate an agent
Personal credentials may get an automation through a demonstration. They also make every resulting action impossible to attribute honestly.
The evaluation set nobody wrote
Without a representative baseline, every prompt or model change is a guess—and every release depends on memory.
When the right answer is a rule, not a model
Some apparent AI opportunities are really unresolved process decisions. Mapping the work first often reveals a simpler and more dependable intervention.