Now
Currently
Machine learning engineer at Sandvik Coromant in Trondheim, working on anomaly detection for metal cutting. The models are LSTM autoencoders over multi-sensor streams off a machine tool, and the hard constraint is not accuracy but time: a detection is only useful if it arrives while the cut is still happening, which in practice means inference in under sixteen milliseconds on hardware already busy doing something else.
Most of the interesting work is on either side of the model — CUDA training pipelines, ONNX on embedded targets, and the C#/.NET and TypeScript plumbing that turns a notebook result into something a product team can ship. Enough of it is full-stack that I built this site in Blazor partly to keep that muscle warm.
Before this, a PhD at NTNU with Jo Eidsvik on adaptive sampling for autonomous underwater vehicles: Gaussian random fields over the ocean, value-of-information criteria for deciding where to sample next, and long-horizon path planning that has to respect obstacles and a battery. It ended with a planner running onboard during a fjord trial, which is a different kind of nervous from waiting on a test suite.
Thinking about
The seam between those two jobs is where I would like to spend more time. Adaptive sampling in the literature usually assumes the information criterion is free to evaluate; in production nothing is free. Concretely, the questions I find myself returning to:
- What happens to a value-of-information criterion when it must be computed inside a hard latency budget — and does the resulting decision survive the approximation?
- Informative path planning where the information term itself is a learned surrogate, and the reality gap belongs to the information, not just the dynamics or the perception.
- Sequential change-point detection and value of information as the same problem in different notation — when to acquire, where, and at what cost.
Always glad to talk to people working on any of that, in academia or industry.
Publications
-
3-D Adaptive AUV Sampling for Classification of Water Masses
Ge, Y., Eidsvik, J., & Mo-Bjørkelund, T.
IEEE Journal of Oceanic Engineering, 2023 · doi · pdf -
Long-Horizon Informative Path Planning with Obstacles and Time Constraints
Ge, Y., Olaisen, A. J. H., Eidsvik, J., Jain, R. P., & Johansen, T. A.
IFAC-PapersOnLine 55(31), CAMS 2022 · doi · pdf
Earlier work on information-driven sampling is in the information-gain note (pdf).