How We Reduced Defect Escape Rate from 4% to 0.3% with Computer Vision

The Problem

The client is a mid-size Czech manufacturer of precision automotive components, producing 12,000+ parts daily across two production lines. Quality inspection was entirely manual — three full-time inspectors visually checked each component for surface defects, dimensional issues, and assembly errors.

The problems were compounding:

• 4% defect escape rate — defective parts reaching customers, causing warranty claims
• Bottleneck during peak periods — inspectors couldn't keep up with production speed
• Inconsistency — inspection quality varied by shift, time of day, and individual inspector
• Cost — three full-time inspectors plus rework costs totaling €200K+ annually

What We Built

We designed a custom convolutional neural network (CNN) trained on 50,000 labeled images of good and defective parts. The system integrates directly with the existing conveyor belt and triggers automatic rejection of defective components.

Technical approach

• Data collection: 2 weeks of capturing images from production line with industrial cameras
• Labeling: Worked with the client's QA team to label defect categories (cracks, burrs, dimensional drift, contamination)
• Model architecture: Custom lightweight CNN optimized for edge deployment — 15ms inference time per part
• Edge deployment: NVIDIA Jetson module mounted directly on the conveyor, no cloud dependency
• Integration: REST API connection to existing MES (Manufacturing Execution System)

Timeline

Weeks 1-2: On-site assessment, camera setup, data collection
Weeks 3-6: Model training, iteration with QA team, accuracy optimization
Weeks 7-8: Edge deployment, conveyor integration, MES API
Weeks 9-11: Production pilot, side-by-side with human inspectors, fine-tuning

Results (6 months post-deployment)

• Defect escape rate: 0.3% (down from 4%) — 13x improvement
• Inspection speed: 3x faster than human inspectors
• Consistency: 24/7 — no variation by shift or fatigue
• ROI: €180K annual savings on a €95K project — payback in 6 months
• Inspectors redeployed to complex QA tasks that require human judgment

Key Takeaway

This wasn't a research project. It was an 11-week engagement that went from "we have a quality problem" to "the system is live and saving money." The client's QA team was involved from day one — they labeled the training data and validated the results. That's why it worked.

If you're inspecting products manually and your defect rate is above 1%, there's a good chance computer vision can help. The question is whether your data and process make it feasible — that's what we figure out in the first two weeks.