Vision AI Quality Control
A computer vision inspection system that detects microscopic defects at 99.7% accuracy on a high-speed production line, replacing manual QC for a leading electronics manufacturer.
Overview
Status: client project. Client details are anonymised under NDA. This page documents the engineering approach and measurable outcomes.
A tier-1 electronics manufacturer was shipping 0.8% defective units despite employing 120 human inspectors across three shifts. Defects ranged from hairline solder cracks invisible to the naked eye to subtle color-shift anomalies on display panels. We designed and deployed a multi-camera edge-AI inspection system that runs inference directly on the production line at 460 units per minute — three times faster than human inspection with an order-of-magnitude improvement in catch rate.
Problem
Human inspectors caught only 91% of defects, and fatigue degraded accuracy by up to 18% during night shifts. The cost of a single escaped defect — including warranty, recall logistics and brand damage — was estimated at $4,200 per unit. The client needed a system that could match line speed of 460 UPM without adding latency or requiring a production shutdown to install.
Approach
We trained an ensemble of convolutional and vision-transformer models on 2.4 million labeled images spanning 37 defect categories, augmented with synthetic data for rare failure modes. Inference runs on NVIDIA Jetson edge clusters mounted directly on the line, connected via high-speed GigE Vision cameras with telecentric lenses. An active-learning loop continuously refines the model as new defect types emerge, with human-in-the-loop verification for edge cases that fall below the confidence threshold.
Result
Detection accuracy reached 99.7% within six weeks of deployment, with a false positive rate under 0.1%. Escaped defect rate dropped from 0.8% to 0.02%, saving the client an estimated $11.6M annually in warranty and recall costs. The system has been rolled out to four additional production facilities.
“The inspection system matched line speed on day one and never needed a shutdown to install. Our first audit after go-live found nothing — which had never happened before.”
Daniel Okafor
Director of Manufacturing Engineering · Electronics manufacturer