Intelligent Robot Vision Monitoring System
Production-deployed computer vision monitoring system with ML-driven anomaly detection, automated clip capture, and MSSQL event logging — built under real manufacturing constraints at Magna International.
Problem
Manufacturing cells had no visibility into runtime behavior — failures were discovered late and investigated manually.
Manual review was slow, inconsistent, and produced no reusable data.
Any monitoring solution had to run without disrupting cycle time on live production equipment.
Solution
Built a real-time inference pipeline for object detection and robot state classification running on live cells.
Designed an anomaly detection and clip-capture engine that automatically saves video evidence on trigger.
Structured MSSQL event logging with a schema purpose-built for dashboards and post-incident queries.
Key Engineering Decisions
Chose unsupervised anomaly detection to work around the absence of labeled failure data in a production environment.
Decoupled real-time inference from logging and I/O to prevent any single slow path from blocking detection.
Profiled and rewrote hot paths in C++ and added targeted SQL indexes after identifying the specific bottlenecks.
Results
Achieved 98% detection accuracy on live manufacturing cells — deployed with unsupervised methods and limited labeled data.
Reduced processing latency from seconds to milliseconds through C++ hot-path rewrites and SQL index tuning.
Enabled searchable event history and auto-captured clips, cutting post-incident investigation time significantly.