All case studies
Manufacturing
Real-time vision QC on a packaging line.
Caught 99.2% of defective units at line speed (38 fps).
Computer visionEdge inferenceYOLO
Outcomes
By the numbers.
99.2%Defect recall
38 fpsThroughput
0.6%False-flag rate
Challenge
The problem we were brought in to solve.
Defects in pre-printed packaging were caught downstream after the products were already shipped. The team needed a vision system that ran at line speed on commodity hardware.
Approach
How we built it.
- Collected and labeled 14,000 images covering 7 defect classes.
- Trained a custom YOLO variant with class-weighted loss for the rare defects.
- Optimized inference with ONNX + TensorRT to hit 38 fps on a single GPU.
- Deployed with a fail-safe rule: any low-confidence frame routes to a human review queue.
Stack
The technology we used.
PyTorch
Ultralytics YOLO
ONNX
TensorRT
Postgres
Related services
Have a similar problem to solve?
Tell us about your project. We respond within one business day.