From Human Eye to AI Eye: How an Auto Parts Factory Went from 85% to 98% Yield Rate
A small auto parts manufacturer in Ningbo, producing brake discs for a joint-venture automaker, faced a classic crisis: 15% defect escape rate in manual visual inspection, leading to 1,800 defective units slipping through monthly. Customer complaints culminated in a 4-hour OEM line stoppage with a penalty exceeding the quarter's profit.
Five inspectors, three shifts, 800 brake discs per person per shift. Human attention fatigue pushed defect escape from 8% (morning) to 22% (night).
The Choice
Three options: hire 5 more inspectors (+¥400K/year, 8% escape), buy German optical inspection equipment (¥1.2M, 6-month lead time, poor model change flexibility), or deploy an AI vision system on KAIHE's cloud gateway with open-source vision models (¥150K, 30-day deployment, self-serviceable).
The factory chose AI. Not for price — for flexibility. Brake disc models change 2-3 times yearly. Traditional equipment requires vendor engineer visits for recalibration. AI systems need only re-annotated samples — managed by in-house QA staff.
30-Day Deployment
Days 1-5: 5,000 annotated defect images (cracks, pores, scratches, dimensional deviations) from 3 industrial cameras.
Days 6-15: Open-source vision model trained 15 epochs. Validation accuracy hit 99% at epoch 12, 99.7% at epoch 15. Stress test: 1 missed out of 100 deliberately defective samples (human would have missed it too).
Days 16-25: KAIHE gateway deployed on production-line industrial PC. Automatic rejection pusher integrated — AI-flagged parts route to secondary manual inspection.
Days 26-30: Dual-track validation. Day 28: AI false-rejection rate <0.3%, zero missed defects. Three of five inspectors reassigned to higher-value process quality roles.
Results
| Metric | Before | After | Change |
|---|---|---|---|
| Yield rate | 85% | 98.2% | +13.2pp |
| Defect escape | 15% | 0.3% | -98% |
| Inspection labor | 15 shifts | 6 shifts | -60% |
| Customer complaints | 3.2/month | 0 (2 months post) | -100% |
The ¥150K investment paid back in two months on saved penalties and rework costs alone.
The real lesson: in 2026, SMEs don't need to build AI from scratch. They stand on existing open-source models and off-the-shelf cloud gateways, focusing solely on the last mile — turning their domain data into AI-comprehensible training samples.