03 · MODEL WORKSHOP
VIAI VisionLab
Train on good parts. Release with confidence.
Import images, train on good parts with a five-step wizard, measure the model on parts it never saw, and send it to every appliance after four release checks.
5 steps
Training wizard, no code
4
Release checks per model
2–4
Runs compared side by side
SHA-256
Every file verified on arrival
PROBLEM → SOLUTION
What it fixes on your line.
PROBLEM
Training AI needs a data-science team.
SOLUTION
Five-step wizard
Task and data, preparation, model, compute, review. No code.
PROBLEM
You never have enough defect images.
SOLUTION
Good-parts training
Anomaly models learn from good parts. Synthetic scratches, dents and cracks fill the gaps.
PROBLEM
A bad model could reach the line.
SOLUTION
Four release checks
Data quality, accuracy on unseen images, export parity and a known-good sample.
PROBLEM
No record of who approved a model change.
SOLUTION
Model cards
Every release records its approver and check results.
PROBLEM
Model files get corrupted or swapped in transfer.
SOLUTION
SHA-256 verification
Every file is checked when it arrives on the appliance.
HOW IT WORKS
From a folder of good parts to a released model.
01
TRAIN
Train without code
Task and data, preparation, model, compute, review. Progress and learning curves live.
02
EVALUATE
Measure on unseen parts
See how good and defective parts separate. Move thresholds and see the trade-off.
03
RELEASE
Release with checks
Four checks, an approver and a model card, then sync or bundle to every appliance.
RELEASE CHECKS
Nothing reaches the line unchecked.
The training data is sound
Resolution, blur, duplicates and class counts.
Accurate on unseen images
AUROC, F1 and false-positive rate.
The exported file matches
Same answers as the model, fast enough.
A known-good sample passes
Catches a model that drifted into rejecting everything.
MODELS
PatchCore
EfficientAD
EfficientNet-B0 classifier
FEATURES
What you get.
01
Trained on good parts
PatchCore and EfficientAD anomaly models react to defects nobody has seen yet.
02
Defect classifiers
EfficientNet-B0 from pretrained weights, from scratch or from an earlier run.
03
Synthetic defects
Scratch, dent, crack, stain, pit and chip, generated for training.
04
Heat-map gallery
Where the heat lands on every image, with ground truth.
05
Release checks
Data quality, accuracy on unseen images, export parity, known-good sample.
06
Model cards
Each release records its approver and check results.
SCREENSHOTS
See it in use.
Anomaly score distribution with review and reject thresholds
Defect classifier: confusion matrix and per-class results
USE CASES
Where plants use it.
QA ENGINEER
New product introduction
Stand up a model for a new SKU before the line ramps.
QA MANAGER
Controlled model changes
Every release is checked, approved and recorded.
ENGINEERING
Improving a model
Retrain from an earlier run and compare up to four runs.