Prepare images
Pin the image set and labels; hold out a labelled evaluation set per defect class.
Industry blueprint
Visual inspection
Manufacturing, packaging and electronics. New defect types keep appearing; retraining must not degrade silently.
Labelling images is expensive and new defect classes emerge over time. Continuous training without fresh labels degrades silently. This design gates every retrained model on held-out images before it reaches the line.
Illustrative pipeline
Train a candidate
Run the partner's vision model on your GPUs inside the budget.
Compare per defect class
Score the candidate against the current model on the held-out images.
Quality approves
Go live only after a quality engineer reviews the comparison.
What the record keeps
The image set and labels, the candidate, per-class scores, the spend, and the approver.
Partner brings
- Camera data and labelled defects
- Vision model
- Connector to the inspection station
Cairn brings
- Pinned image sets and held-out evaluation
- GPU budget
- Per-class comparison and approval
- A record of every model on the line
Business value to test
Escape rate, false-reject rate and time to cover a new defect class.
Keep the decision explicit
A model reaches the line only after a recorded approval.
Close the feedback loop
Images flagged by inspectors become labels for the next candidate.
Reference design, not a shipped domain pack. A partner's vision model runs through the engines Cairn supports.
See how failure can trigger a reviewed improvement cycle →Catch new defects without silent regressions.
Talk about your cameras, defect classes and labelling process.