Thinking
Essays from the team on building AI that holds up in production. Slow cadence on purpose, we publish when we have something to say.
Agentic demos look great. They lose to HITL the moment a decision has a dollar amount attached. A practical taxonomy of where the loop should close, and where it absolutely should not.
The fastest way to wreck an established operation is to migrate the ERP on the way to "becoming AI-native." We build over the system of record, not under it. Here is the math.
Where the human actually sits in an HITL system, what they see, and how to design the queue so a single reviewer can clear hundreds of decisions an hour without losing the plot. The interaction-design rules we keep coming back to.
Two weeks, fixed-fee, against your real systems and real operating people. What we ask, what we deliver at the end, and what makes us walk away. Doubles as a procurement document.