.png){kind=logo}
Blog
Notes on making AI answer an inspection question
What we are learning building inspection AI that a utility can act on and audit. Method and argument, not product announcements.
- Trust and evaluation
Evidence beats confidence
A probability tells you how sure a model is. It does not tell an inspector what to look at, and it cannot be corrected. What field trust actually requires is a chain of evidence attached to every answer.
18 August 2026 / 5 min
- Operating model
The unit of progress is a question, not a model
Model-centric roadmaps produce steady technical progress and unpredictable business progress. Organising the work around individual inspection questions fixes the mismatch, and changes how a team is staffed.
18 August 2026 / 5 min