See, learn, decide, act.
A closed loop, not a one-off detector. The system learns your product from your own images, scores every part in under a second, and gets sharper every time a person corrects it.
See
A camera captures every part on the line and every frame of the floor, at line speed — the hardware you already have is enough to start.
Learn
A model trained on your own images knows what "good" looks like for your product. No two factories are the same, so no two models should be.
Decide
A verdict in under a second: GOOD, BAD, or UNSURE when the call is genuinely too close — so a person decides instead of a coin flip.
Act
Reject the part, raise a safety event, or feed the numbers straight to your dashboard. Every operator correction trains the next model.
A model of your product, not a stock dataset.
Your parts, your defects
Operators label images in your own vocabulary — scratch, dent, contamination — each named, coloured and carrying its own verdict. That becomes a training set in one click, and a fresh model trains in minutes on our cloud GPUs.
It says when it isn't sure
An anomaly line can still show a plausible score without a perfect model — but a verdict near the threshold is flagged UNSURE and sent to a person, instead of guessing. You see exactly why a part was rejected, with the region marked on the image.
Runs on the edge.
Inference runs next to the camera, so your images never leave the line and a verdict lands in milliseconds — no round-trip to a cloud, no bandwidth bill, no data leaving the building. Faces are blurred at playback, and you set how long anything is kept.
See it on your parts.
Bring a handful of photos of your product. We'll train a model and show you the verdicts — usually within the first call.