I'm Yaolin. I work on the part of machine learning that has to decide what to measure next — first through a PhD at NTNU, where that meant an autonomous underwater vehicle choosing where in a fjord to take its next sample, and now at Sandvik Coromant, where it means anomaly detection on multi-sensor machining data that has to return an answer in under sixteen milliseconds.

The two halves inform each other more than I expected. Academic adaptive sampling tends to stop at simulation or a single field trial; industrial anomaly detection often has no decision-theoretic grounding at all. The question I keep circling is what happens to a value-of-information criterion when it has to be computed inside a hard latency budget — and whether the decision survives the approximation.