Client name, imagery and figures are withheld pending approval. The engagement structure, architecture and technical reasoning shown here are the real template we publish against.
VISUAL INSPECTION
WHEN THE DEFECT
IS RARE.
The critical defect appeared about twice a month.
The line ran fast enough that human inspection sampled rather than checked. Most defect classes were common enough to train on comfortably. One was not — the class that caused warranty claims occurred a handful of times a month across the whole plant, which meant the historical archive held fewer than two hundred confirmed examples, unevenly lit and inconsistently framed.
A model trained on that archive scored well on the common classes and effectively guessed on the one that mattered. Accuracy as a single number hid the failure completely.
Synthetic defects, generated at volume.
The obvious move was to synthesise the rare class — augment the two hundred examples into twenty thousand. Validation scores improved substantially. On the line, performance was worse than the baseline.
The synthesis had reproduced the appearance of the defect but not its context: real occurrences correlated with a particular tooling wear pattern that changed the surface texture around them. The model had learned our generator, not the plant. We threw the approach out in week five and said so in the weekly review.
Stop classifying. Start ranking what a human looks at.
We reframed the problem. Instead of asking the model to decide pass or fail on the rare class, we asked it to rank every part by how unusual it looked — an anomaly score trained mostly on good examples, of which there were millions. The operator's screen showed the twenty most unusual parts per shift.
Every operator judgement on those twenty fed straight back as a label. Within eleven weeks the confirmed archive of the rare class had roughly tripled, and a supervised model for it became viable — trained on real examples the system had surfaced itself.
{{ l }}
HAVE A RARE-EVENT
PROBLEM?
Scarce examples of the thing that matters most is one of the most common shapes we see. There are established ways through it, and none of them start with generating fake data.