AI Quality Inspection for Manufacturing
Defect detection from 87% to 99.2%. $340K saved in year one. Saved the largest retail account.
Client details anonymized under NDA. The work, approach, and results shown here are real. Contact me for references.
Defect Detection
Saved in Year 1
Defect Types Classified
Inspection Speed
The Challenge
What they were dealing with
A consumer electronics manufacturer running two production lines was relying on manual visual inspection for quality control. Four inspectors working in shifts caught about 87% of defects, but the 13% that slipped through resulted in expensive returns, warranty claims, and retailer chargebacks. The company was spending over $400K annually on quality failures, and their largest retail partner had issued a warning about defect rates.
Human inspectors get fatigued after two hours of staring at products on a conveyor belt, and miss rates spike toward the end of each shift
Some defects like hairline cracks and micro scratches are nearly invisible to the naked eye under production lighting
No data on which defect types were most common, making it impossible to fix upstream production issues
The largest retail partner threatened to reduce order volume if defect rates did not improve within six months
Before
87%
Defect Detection
$400K+
Annual Quality Costs
13% miss rate
Inspector Fatigue Gap
Active
Retailer Warning
The Approach
How I solved it
Human inspectors are great at catching obvious defects but terrible at maintaining consistency over an eight hour shift. I measured their detection rate by hour and the data was clear: 95% accuracy in the first two hours, dropping to 78% by hour six. The overall 87% average was hiding a much worse problem at the end of each shift, which is exactly when the highest volume runs happened.
I built a computer vision system that runs alongside the existing production line, inspecting every single unit in real time with cameras positioned at four angles (top, bottom, left, right). The system was trained on 15,000 images of both good and defective products, covering 23 defect categories ranging from cosmetic scratches to hairline structural cracks that are invisible to the naked eye under production lighting. The model runs on edge compute hardware mounted directly on the line for sub 50 millisecond inference, which is fast enough to keep up with the conveyor belt at full speed.
When the system detects a defect, it triggers an automated pneumatic reject mechanism that diverts the unit off the main line before it reaches packaging. Every defect is logged with a high resolution photo, classification, confidence score, timestamp, and station ID. The real bonus was the analytics. Within the first month, the defect data revealed that 40% of cosmetic scratches were happening at one specific station due to a worn conveyor guide rail. Fixing that $200 part reduced the overall defect rate by another 15%. That single insight paid for the entire camera system.
Data Collection and Labeling
Captured and labeled 15,000 images across 23 defect categories using cameras installed at four angles on the production line.
Model Training
Trained a real time object detection model optimized for production speed. Sub 50 millisecond inference to keep up with conveyor belt pace.
Line Integration
Installed camera mounts, lighting rigs, and automated reject mechanisms on both production lines with zero downtime during installation.
Analytics Dashboard
Real time defect tracking dashboard with shift level reporting, defect type breakdown, and trend analysis for upstream process improvement.
The Results
What changed
99.2%
Defect Detection
$340K
Saved in Year 1
23
Defect Types Classified
50ms
Inspection Speed
“We were about to lose our biggest retail account over quality issues. Jahanzaib installed the vision system and within a month our defect rate went from 13% to under 1%. The retailer increased their order volume. That system paid for itself in the first quarter.”
Robert Tanaka
VP Operations, Consumer Electronics Manufacturer
Related Projects
AI Inventory Forecasting for Ecommerce
Built an ML forecasting system that segments 2,000 SKUs into behavioral clusters, predicts demand per product type, and auto generates purchase orders across three warehouses.
94% Prediction Accuracy
AI Contract Review for Law Firm
Built an AI system that reads contracts in any format, classifies 45 clause types, compares each one against the firm's approved playbook, and generates a structured redline summary with risk ratings. Partners now review a one page AI summary instead of a 40 page contract.
15min Avg Review Time
Facing a similar challenge?
Every project starts with a conversation. Tell me what you're dealing with and I'll tell you honestly whether I can help.