Most of our machine-learning work runs on data that came off an instrument we helped design. That changes the problem. There is rarely an abundance of labelled examples, the measurement process has its own structure and artefacts, and a confident wrong answer can be considerably worse than no answer at all.
What we work on
We build models for scientific and sensor data, and we deploy a good proportion of them on devices with a few hundred kilobytes of memory and a power budget measured in milliwatts. Working within those limits tends to produce better engineering: it forces a clear account of what the model is actually for and what the smallest thing is that would do the job.
Often the honest answer is that the job does not need a learned model. A well-understood signal-processing chain that can be reasoned about, tested exhaustively and explained to a regulator is frequently the better outcome. We will say so.
Evaluation is the work
A headline accuracy figure on a convenient test split is close to meaningless for a system that has to operate on new subjects, new hardware revisions and conditions nobody sampled. We spend our effort on the parts that predict real behaviour: how the data was split and why, how performance moves across subgroups and operating conditions, what the model does when its input goes out of distribution, and what it costs when it is wrong.
Calibrated uncertainty and a defined abstention path matter more than a point estimate whenever a person is going to act on the output.
Governance
We document what a model was trained on, what it was evaluated against, the conditions it was not tested under and who is accountable for the decision it supports. For anything touching health, safety or personal data, that record is part of the deliverable rather than an optional extra — and the provenance and licensing of training data is settled before training, not after.
Working with us
If you have a measurement problem that people keep describing as an AI problem, we are happy to help work out which of those it actually is. Get in touch.