Teach a new product in an afternoon.
Your quality team adds new products and new defect types with a few dozen labeled images. No code and no data science team. Every model is tested before it reaches the line.
Images straight from the line
Collect images from the running camera or upload your own boundary samples.
Label in the browser
Draw a box or brush over the defect and pick its type. Good parts need a single click.
Few images needed
30 to 50 good parts and 10 to 20 examples per defect type to start.
Tested before release
Catch rate and false reject rate on held-out images, shown before you deploy.
Versioned models
Every model is versioned and can be rolled back with one click.
Learns from review
Parts your team judges in review become new training examples.
From new part to live model.
Model Trainer at a glance
| Images to start | 30 to 50 good parts, 10 to 20 per defect type |
|---|---|
| Training time | Typically 20 to 60 minutes |
| Labeling tools | Box, polygon, brush, whole-image good or bad |
| Release checks | Catch rate, false reject rate, review rate on held-out images |
| Deployment | Shadow mode, then live, with one-click rollback |
| Available on | 40× and 100× plans |
Asked by quality teams
Do we need data scientists?
No. The trainer is built for quality engineers. We help with your first product during the pilot.
What if a defect is very rare?
Start with the examples you have. The model sends similar-looking parts to review, and each confirmed case strengthens it.
Can one model cover several product variants?
Yes, when variants look alike. Different shapes or finishes usually get their own model, switched by recipe.
Turn the knob to your line speed.
A 5% manual sample checks 5,760 of them. NanoQuantumAI checks all 115,200, and keeps a picture of each one.