How it learns
You place good parts in the fixture and take photos. About 60 is the working number, and the setup wizard takes about 30 minutes. VZSOR stores what normal looks like for that one part, from that one camera position, under that one light.
After that, every new part is compared to the stored picture. Areas that look unlike anything in the good set are marked on a heatmap. The station shows PASS or FAIL, and the operator sees where it looked. The result appears while the operator is still holding the part, so the check fits into the normal motion of the bench.
Why good parts only
A defect library is hard to build. Real defects are rare, and each one looks a little different. A shop that makes ten of a part per week may never collect enough bad examples to teach a system what bad looks like.
Good parts are easy to get. Your line makes them every day. Training on them also means the station can flag a defect type nobody listed in advance, because it is not matching against a list. It is looking for anything out of the ordinary.
The trade is that it flags unusual, not wrong. A harmless mark can look unusual too. That is why a person reviews each flag. A harmless mark can look unusual too, so some good parts will be flagged. A pilot counts false rejects on your parts before you buy anything.
What a flag means
A FAIL from VZSOR means this part looks different from your good parts in a way worth a second look. It does not mean the part is scrap. The operator checks the photo and the heatmap, then holds the part or releases it.
If the operator disagrees with the verdict, they override it and type a reason. The override is saved next to the original result. The AI verdict is never overwritten, so the record shows both. This also gives a quality manager something to review later: how often operators agreed with the station, and how often they did not.
What has been tested so far
VZSOR has been validated on one part: a 2 inch by 2 inch aluminum L-bracket with a mill finish. It was trained to catch dents, scratches and nicks on the visible faces. That is the whole claim. Everything else, including other part shapes and finishes, is untested.
Your parts are different. A paid pilot trains the station on your part and measures it on a test set you help choose, so the result comes from your parts and not ours. The pilot fee is credited toward a station if you buy within 30 days, and the success terms are written down before it starts.
How a shop sets it up
- Pick one part number to start. The model is trained per part.
- Fix the part in a fixture so it sits the same way every time.
- Photograph about 60 good parts from your line.
- Run the setup wizard, about 30 minutes.
- Keep the camera, light and fixture where they are. If one moves, retrain.
Limits, stated plainly
- Validated on one part only: a 2 inch aluminum L-bracket, mill finish.
- Anodized parts and tool marks are not validated. Tool marks may look unusual to the model.
- A flag means unusual, not scrap. Expect some good parts to be flagged.
- Results on your parts are unknown until a pilot measures them.