Telling drivers apart from the way they drive

The problem

A single vehicle was driven across continents in eleven continuous days by a rotating crew. The question put to us was deliberately hard: using only what the car recorded, reconstruct how many distinct drivers there were, and when each of them was driving.

Nobody handed us labels. This is an unsupervised problem — the answer has to come out of the structure of the data itself, and the only way to know whether it is right is to check it afterwards against a log we were not allowed to see first.

What we did

We fused GPS traces, in-car sensor streams and environmental data into a single feature space, then used principal component analysis to compress dozens of correlated measurements into a handful of parameters that actually carried the variation.

Clustering in that reduced space separated the driving signatures. Three groups emerged, with some overlap at the boundaries — which is what honest clustering looks like when two people drive similarly on the same stretch of road.

The result

Validated against the true driver log afterwards, the model identified who was driving with 78% accuracy.

Worth being precise about what that means. It is not a claim that driving style is a fingerprint. It is a demonstration that behavioural structure survives in ordinary vehicle telemetry, and that it can be recovered without labels — which is the same capability that finds unexpected regimes in industrial sensor data.

What we did not build

The original deck proposed using this to price motor insurance individually. We do not put that forward. The defensible applications are the ones where the subject is the road rather than the person — detecting surface degradation, hazardous stretches and changing conditions from ordinary fleet telemetry.