Cutting defect detection time from minutes to under 2 seconds

The Challenge

The manufacturer's existing camera system detected only 60% of solder joint defects on PCB assemblies. Manual inspection added 3 minutes per board, creating a production bottleneck. Defective boards that slipped through caused costly rework and customer returns.

The Solution

A Hungarian electronics manufacturer needed to inspect solder joints on PCB assemblies. Their existing camera system caught about 60% of defects, and manual inspection added 3 minutes per board. We trained a custom vision model on 14,000 labeled images from their production line. It wasn't a plug-and-play solution — we spent two weeks just getting the lighting and camera angles right so the model could work consistently across shifts. Now it runs inline at 1.8 seconds per board with a 94% detection rate.

Technical Approach

We trained a custom CNN on 14,000 labeled images from their production line. Two weeks were spent optimizing lighting rigs and camera angles for consistent results across all shifts. The model was deployed inline using FastAPI with GPU inference, integrated directly into the existing conveyor control system.

Technology Stack