Batch-processes container images: YOLO finds the ID plate, TrOCR reads it, and every code is validated against ISO 6346. Operators review flagged results, correct them, and files auto-rename — with full batch history in SQLite.

Real-time glove compliance monitoring with up to six operator-defined detection zones on live video, wired directly into the production line via bidirectional Modbus TCP PLC communication.

Two models on one live feed: YOLO classifies customers vs. employees at 90%+ accuracy, while RT-DETR spots small, occluded items like cups and bottles at 95%+ — trained on imbalanced data using SMOTE.

Detects and locks onto people in video with stable identity tracking frame to frame — the foundation for counting, dwell-time analytics, and multi-camera re-identification.

YOLO11-pose extracts body keypoints and a lightweight CNN+GRU classifies 1.5-second windows of movement as fall or non-fall. Trained on CAUCAFall with a subject-level split to avoid leakage: 10/10 falls detected with no false-alarm clips on held-out subjects (F1 0.96, ~0.3 s delay). Also tested on AI-generated CCTV-style clips; real-world validation is next.