Vossberg Automotive · Computer vision

Eyes that never blink on the line.

Vossberg Automotive inspects thousands of components an hour. Manual spot-checks let defects slip through to costly downstream stages. Pulse watches every part, in real time, on the cameras the factory already owned.

ClientVossberg Automotive
IndustryManufacturing
Year2024
Timeline16 weeks
ServicesComputer vision · Edge AI
Pulse
01 — The challenge

Spot checks can't see everything.

Quality control sampled roughly one part in twenty. Defects discovered at final assembly cost up to fifty times more than at the station where they originated — and some reached customers.

The line runs at takt times that leave under 40 milliseconds per frame for a decision, on a factory floor where new hardware approval takes a year. The system had to run on existing cameras and industrial PCs.

02 — The solution

Inference at the edge, answers in milliseconds.

We trained defect-detection models on the plant's own imagery, compressed them for edge inference, and deployed them to the industrial PCs already mounted on the line. Pulse flags a defect before the part leaves the station and signals the PLC to divert it.

A dashboard tracks defect classes over time, turning quality data the factory never had into process improvements upstream.

What we built

  • Defect models trained on plant imagery, tuned per station
  • Edge inference under 40 ms on existing industrial PCs
  • Direct PLC integration for automatic part diversion
  • Continuous learning loop with operator-confirmed labels
  • Quality dashboard linking defect classes to process causes
03 — The results

Every part inspected, none slowed down.

Pulse now inspects 100 % of parts at full line speed. Detection reached 99.2 % with a false-positive rate operators actually tolerate, escaped defects fell by two thirds, and the system paid for itself in seven months.

99.2%detection rate
<40msper frame
−67%escaped defects
7 mopayback period
“We expected a research project. We got a production system that our line operators trust — running on hardware we already owned.”
Dieter VossbergHead of Production · Vossberg Automotive
04 — Tech stack
PythonPyTorchTensorRTOpenCVNVIDIA JetsonC++OPC UAGrafana