Containerized Automated Quality Control Pipeline

The Challenge

Manual visual inspection on manufacturing lines is the primary bottleneck for throughput, subject to human fatigue and inconsistency. The objective was to prototype a computer vision system capable of identifying product presence and defects in real-time, with a strict requirement for portability across different production servers.

I engineered a containerized object detection pipeline capable of processing video feeds at 30+ FPS. Rather than training a custom architecture from scratch, I leveraged transfer learning on the YOLOv8 (You Only Look Once) architecture to ensure low-latency inference suitable for edge deployment.

Inference Engine: Utilized Ultralytics YOLOv8 for single-shot object detection. Tuned confidence thresholds (Confidence > 0.5) to balance the trade-off between Recall (catching all objects) and Precision (minimizing false stops).

Video Processing: Implemented an OpenCV wrapper to handle video stream ingestion, frame buffer management, and bounding-box annotation overlays.

MLOps & Deployment: Dockerized the entire application. I wrote a multi-stage Dockerfile to handle complex dependencies (like libGL for OpenCV), ensuring the model runs identically on a local development laptop and a cloud-based inference server.

Core: Python 3.9, Docker

Computer Vision: YOLOv8 (Deep Learning), OpenCV (Image Processing)

Environment: Linux (Debian-based container)

Delivered a portable, "write once, run anywhere" inference system. By containerizing the solution, I eliminated the "it works on my machine" dependency issue common in Python CV projects, proving the system is ready for integration into a larger microservices architecture.