Human inspectors miss ~8% of cosmetic defects on a fast belt. Rework cost is ~$120k/mo. Throughput is capped by inspector hours.
The AI approach
Industrial camera over the belt; object-detection model spots defects per frame; classifier identifies type + severity; line PLC reroutes or rejects.
The outcome
~99.2% recall
Throughput +18%. Inspector reviews only ambiguous cases — a calibration role, not the front line.
Try itInput → Process → Output
Input — Camera frame
LINE 04 · CAM-AF#2026-05-12T14:08
scratch · 0.91
Process — AI pipeline
1Capture frame from cameraSensorSymbolic
2Detect defects (YOLO)VisionPredictive
3Classify type + severityPredictPredictive
4Write note · route to rejectWriteGenerative
Output — Defect record
Click Run demo to detect, classify, and route the unit.
DEFECT EVENTREJECT · auto
VerdictREWORK ▸
Unit IDU-2026-7714201
Defect typeSurface scratch
Confidence0.91
Locationlabel · upper-left
Severity
6.2 · MED
Routed
Belt diverter set to lane 2 (rework). Note added: "Scratch ~12mm parallel to label edge — check upstream guide roller." MES event QC-EV-44128 created.
Three AI types in this use case
SymbolicFrame trigger from line encoder, belt-divert PLC commands, severity-to-action thresholds, MES event schema.
PredictiveObject detection (YOLOv8) + defect-type classifier (ResNet) — the engine that actually sees the scratch.
GenerativeLLM writes the human-readable defect note so the maintenance engineer can look upstream without watching footage.