Case Study — Q4 2024

Computer vision defect detection for manufacturing production lines

Architected an end-to-end computer vision inspection system across 3 production lines — designing the data pipeline, model serving infrastructure, and edge deployment strategy that catches 99.1% of defects at 30fps, saving €0.8M annually in scrap and rework.

Role

Solution Architect

Client

Manufacturing

Duration

6 months

Team

4 people

Results after 90 days in production

99.1%
Defect Catch Rate
Up from 87% manual
42%
Downtime Reduction
Predictive scheduling
€0.8M
Annual Savings
Scrap + rework costs
30fps
Processing Speed
Real-time 4K feeds

Interactive Detection Pipeline

How the visual inspection system works

Click a sample to simulate inspection

📷 4K Feed
CNN EfficientNet-B4

The Problem

Quality control that couldn’t keep up with production

The client operated 3 high-speed production lines outputting 2,400 parts per hour. Manual visual inspection caught only 87% of defects — dropping to 79% on night shifts. Each escaped defect cost an average of €3,200 in warranty claims. But the deeper problem was architectural: there was no scalable system in place to integrate vision AI into the existing factory infrastructure.

Legacy PLCs, proprietary camera protocols, air-gapped networks, and zero MLOps maturity meant this wasn’t just a model problem — it was a full-stack systems integration challenge. The client had already failed one pilot with a vendor who delivered a high-accuracy model but no viable path to production deployment. They needed a solution architect who could design the entire pipeline end to end — from camera placement to model serving to operator dashboards — and make it work within their existing OT environment.

Architecture Decisions

Key architecture decisions I drove

Execution

6-month delivery timeline

Technology

Stack & tools