Automated Factory Inspection

A global manufacturing operation was running automated production lines with a visual quality-control step that depended on manual review.

Human inspectors could not keep pace with line throughput. Defect review became a bottleneck in a live production workflow, and every uncertain case either delayed the line or escaped as a quality risk.

• Decisions had to happen on the live camera stream, inside the production line's timing budget - not in a batch process after the fact.
• Both error types were costly: false positives rejected sellable product, false negatives let defects escape.
• The system had to run continuously in a production environment, not as a demo or a scheduled experiment.
• Inference had to run at the edge, close to the cameras, inside the facility's existing infrastructure.

The sanitized architecture follows the path the frames take through the system:

Vision models for defect detection were trained and evaluated against production imagery, then optimized for edge deployment with TensorRT.

• RTSP ingestion with zero-copy GPU transport - no per-frame CPU copies on the hot path.
• TensorRT-optimized inference: 11.4 ms kernel latency, under 15 ms end-to-end pipeline latency.
• Human-in-the-loop review with adjustable thresholds, so precision could be tuned per defect class without retraining.
• Monitoring and observability on the deployed service, so drift and degradation are visible before they become escapes.

The system runs as production infrastructure: 97%+ precision sustained in live operation, with edge inference latency under 15 ms.

Human review effort is concentrated on the cases the model flags as uncertain, instead of every unit on the line.

Client identity and selected implementation details are withheld under confidentiality obligations.