Catch the fault early, before it voids a six-figure run.

Physical vehicle testing runs across multiple days, with a full crew and expensive track time. A single sensor fault missed in the early laps can invalidate the entire run, a six-figure write-off. We built the model that catches those faults up front, before the costly run is ever committed.

The challenge

Carmakers screen for sensor faults in a handful of manual laps before the main programme. It's slow and prone to human error, and anything that slips through only surfaces once the multi-day run is already underway, by which point the data is invalid and the run is wasted. Across hundreds of noisy channels, reliably spotting the one that's drifting is beyond what a person can watch.

What we built

A reconstruction-based LSTM, built in PyTorch, modelling 100+ sensor channels. It learns each vehicle's nominal baseline from clean driving data. The early screening laps are captured and processed through the model, which flags any channel drifting from that baseline, so faults surface before the multi-day run is committed, not after it has already been spoiled.

The results

Faults surface up front instead of invalidating a run days in, and the model became a standard product within the platform we sold, adopted across multiple carmakers. The architecture then carried into battery testing, a different domain entirely. Not a lift-and-shift, but enough of the design transferred that we stood up the new use case far faster than building from scratch, which is the real leverage: prove the approach once, then redeploy it across domains.

Our role

Delivered before re:rate. We led the build, a team of two to three data scientists, setting the modelling approach and architecture and reviewing the work against the client's real problem. We owned the technical client relationship, gathering the initial requirements, bootstrapping the project and running the feedback loop, and drove it from a first prototype in months to a full production system inside a year.