Case study · 2025 · Manufacturing
Vision Inspection on a High-Speed Packaging Line
A consumer-goods manufacturer was catching label and seal defects the only way it could: sampling by eye, at a line speed no human can fully watch. We built camera-based inspection that sees every unit — detection models running on an edge GPU beside the line, an operator review station, and a retraining loop that turns each operator decision into training data.
Screens are illustrative recreations — client interfaces stay confidential.
The challenge
At 96 units a minute, manual sampling inspected a fraction of production; the rest shipped on faith. Escaped defects surfaced as retailer chargebacks and, occasionally, as a pallet returned in full.
The factory floor sets its own rules: variable lighting, product changeovers several times a shift, and no reliable path to ship frames to the cloud — inference had to live at the line and adapt to new SKUs without a modeling project each time.
What we built
Capture built for the line
Machine-vision cameras with controlled strobe lighting at two stations, triggered per unit by the line encoder — consistent frames at full speed, unaffected by ambient factory light.
Edge inference in the millisecond budget
Compact detection models quantized and served with ONNX Runtime on an industrial GPU box at the line, classifying seal, label, and print defects inside the per-unit time budget with automatic reject signaling.
Operators as the training loop
Flagged units land in a review station where operators confirm or correct in one tap; those labels stream into scheduled retraining, so accuracy climbs with production instead of decaying.
Changeover-friendly by design
New SKUs onboard through a guided capture session of golden samples; the pipeline fine-tunes and validates against a held-out set before the model is allowed onto the line.
The results
- 99.2%
- Defect recall
- −84%
- Escaped defects
- 96 / min
- Units inspected
seeded-defect audits at full line speed
41 → 6 per 100k units over six months
every unit, both packaging lines
From the manufacturer's monthly outbound quality audits; deployment completed end of October.
View the data as a table
| Escaped defects per 100k | |
|---|---|
| Nov | 41 |
| Dec | 33 |
| Jan | 22 |
| Feb | 15 |
| Mar | 9 |
| Apr | 6 |
- Escaped defects fell from 41 to 6 per 100k units in six months of monthly audits.
- Recall on seeded-defect audits reached 99.2% at line speed, with false rejects tuned below the level that annoys operators into distrust.
- Manual re-inspection labor dropped 72%, redeployed to changeover and quality-improvement work.
Client identities stay confidential; figures are rounded from end-of-engagement delivery reporting.
Stack & expertise
- Python
- PyTorch
- ONNX Runtime
- Edge GPU (Jetson-class)
- Docker
- React
Industries
Technologies
Grounded Research Assistant for a Law Practice
A permission-aware retrieval assistant over 8,000 matter documents — every answer cited, every source access-controlled — that cut research time by two-thirds.
B2B Commerce Replatform for a Building-Materials Distributor
A headless storefront wired live into the ERP — contract pricing, branch stock, quote-to-order — that took online order share from 8% to 27% in nine months.
Route Intelligence for a Regional Courier Network
Dispatch platform with learned travel times and constraint-aware route building that cut fleet distance per stop by 16% without adding vehicles.