Field Notes
Why agentic AI fails in production
The demo always works. Then real volume, real edge cases, and silent drift show up. Here is where agents actually break, why the model is rarely the culprit, and what we watch for.
Operate, don't hand off
Shipping a model is the easy 20 percent. The value is in staying on the controls: monitoring, tuning, and accountability for the outcome.
Keeping PHI in your own cloud
For healthcare, the architecture matters more than any badge. Why agents that run inside your tenant beat sending patient data to someone else's.
The 5-day Diagnostic, explained
What happens in five days, what you walk away with, and why it is the honest front door.
What a dental front desk taught us
Scheduling, recalls, insurance questions: the workflows that quietly broke until we measured them.
Measuring agents you can't fully script
When an agent takes actions, accuracy is not enough. How we set guardrails and track real outcomes.
Plain, from production, no hype.
Field notes on building, measuring, and operating agentic AI in the real world: the patterns that hold up under live traffic, the failure modes that never make the keynote, and the tradeoffs you only learn by running the systems yourself. Written from production, for the people doing the work.