We Replaced 80% of Manual QA with Computer Vision
A real deployment on a 24/7 manufacturing line. Edge inference, defect grading, and the accuracy numbers after 6 months.
A Tier-1 auto parts manufacturer came to us with a problem: their manual QA couldn't keep up with the production line. Human inspectors were missing defects, grading was inconsistent across shifts, and there was no data on what was going wrong.
The setup
We deployed camera stations at three critical points on the press line. Each station runs a custom YOLO-based model on an edge GPU (NVIDIA Jetson), doing inference in real-time as parts move down the line.
The system grades each part into three categories: pass, marginal (needs human review), and reject. Marginal parts get flagged for a human inspector — this is the sweet spot where AI and humans work together.
What we learned about edge deployment
- Lighting is everything. We spent more time on lighting fixtures than on model architecture. Consistent, diffused lighting eliminated 60% of false positives.
- INT8 quantization works. We quantized the model from FP32 to INT8 with less than 1% accuracy drop. Inference time went from 45ms to 12ms.
- The camera matters more than the model. A $2,000 industrial camera with proper optics outperformed a $200 webcam with a state-of-the-art model.
The numbers after 6 months
- 4x faster QA throughput
- -62% scrap rate (catching defects earlier means less wasted material)
- 99.2% detection accuracy on known defect types
- 3 new defect types discovered by the system that humans were missing
Root-cause feedback
The real value wasn't just catching defects — it was the data. By logging every defect with timestamp, position, and type, we built a dashboard that shows operators which press settings correlate with which defects. Preventive, not just detective.