VisionQC — Manufacturing Defect Detection

VisionQC was commissioned by a precision components manufacturer that needed automated quality inspection without slowing production throughput. The solution had to integrate with existing MES workflows and provide operators with actionable defect classifications — not just pass/fail signals.

Manual inspection created bottlenecks and inconsistent defect logging. False negatives risked warranty claims; false positives caused unnecessary line stops. The factory environment introduced variable lighting, motion blur, and reflective surfaces that broke earlier OpenCV rule-based approaches.

I trained a CNN ensemble on 180k labeled production images with heavy augmentation for lighting variance. Models were containerized with Docker and deployed to edge GPUs co-located with camera rigs. Inference results streamed to the factory MES via a secure MQTT bridge with audit logs for compliance.

- 96% defect detection accuracy on validation holdout - Inspection cycle time reduced from 4.2s to 0.6s per unit - Warranty-related returns down 11% in the first quarter post-rollout - Operators received localized defect heatmaps for faster root-cause analysis