Quality Vision — catch the defect on the finished part

A production-ready computer vision pipeline that classifies casting parts as defective or non-defective, with a visual explanation of where the defect is using Grad-CAM.

⚠️ This is an independent portfolio project built on public data — not a paid client engagement. Results are shown as-is, including known limitations.

Key results

Model	EfficientNet-B0 (fine-tuned) + Grad-CAM
Test accuracy	99.72%
Precision / False positives	100.00% / 0
Est. annual savings (10,000 parts/day line)	~$150K vs. ~$15K one-time cost

How it works

Each part image is resized and normalized, then classified by a fine-tuned EfficientNet-B0. Grad-CAM generates a heatmap showing exactly which region of the part triggered the defect prediction, so an operator can trust and verify every decision instead of treating the model as a black box.

Have a similar problem in your operation?

Talk to us about your project