Finding hidden defects in test data before units ship

This work was led by GSWE's founder as senior AI engineer at an automotive electronics manufacturer. Client details are confidential.

The problem

Some units pass every test limit and still fail in the field months later. Each one is a warranty return, and there are too few failures to train an ordinary model.

What we built

• Data pipelines that collect hundreds of test measurements per unit
• Dimensionality reduction (PCA) and variational autoencoders that learn what normal production looks like, with no defect labels needed
• An anomaly score for every unit, with outliers held for re-test before they ship
• Live dashboards so engineers can see drifts as they happen

The result

Units that look normal one test at a time but abnormal as a whole are caught at the factory instead of by the customer.