Day 4 of my daily creative challenge. Today: Harmattan for infrastructure inspection. Pipelines, transmission lines, refineries. A drone flies the corridor and photographs every metre, thousands of frames per kilometre. No human can review all of that, and none should have to. So a detection agent takes the first pass. It flags a coating loss on a weld joint, scores its confidence, and says why: "this patch grew 40% since the last pass and sits on a weld joint." A certified engineer then confirms or rejects each finding. Nothing reaches the client report unless a person has signed it. The part I find most interesting as a designer is that the relationship runs both ways. Every confirm and reject is training data with the engineer's name on it. Over six months the agent's precision climbs and its false alarms fall, because it learns from the reviewer, not the other way round. Known test findings are slipped into the queue to keep both sides honest. What comes out is a report the client can act on, and an asset history that tracks a fault on the weld itself, pass after pass, until a repair closes it. This is a fictional product, designed end to end. Screens below. One of the pros of this kind of product is to enable oil refineries, rigs and other infrastructure to detect defects, worn out equipment, leakages before they happen, thus saving them the cost for repairs and loss of valuable resources flowing through those pipes. I design interfaces for AI agents. If you're building one, my inbox is open.