Descriptions:
A presentation at AI Dev 26 in San Francisco demonstrates Sentinel, a production video anomaly detection system built on a hybrid edge-cloud architecture. The system is designed to solve a real operational problem: surveillance operators managing dozens of cameras cannot manually review continuous footage, and sending all video to the cloud for analysis is both too slow and too expensive at scale.
The technical architecture uses NVIDIA Jetson edge hardware running a lightweight EfficientNet-B0 model to generate video embeddings locally, combined with Quadrant Edge’s embedded vector index and dual-shard KNN search to identify clips that are dissimilar from a learned behavioral baseline. Clips that exceed a dissimilarity threshold are escalated to 12 Labs in the cloud, where more powerful video embedding models perform deeper classification. Quadrant Cloud continuously syncs updated baseline vectors back to edge devices, allowing the system to adapt as environments change — a new room configuration or event space automatically shifts the normal baseline.
Benchmarked on a University of California public anomaly detection dataset with student-simulated events (fighting, running, irregular behavior), the system achieves 0.96 AUC-ROC and 94% recall, producing approximately two false positives per hour on 30-second clips. The edge-first design reduces the total video bandwidth sent to the cloud to just 10% of raw footage volume, making the architecture economically viable for large-scale deployments. GPU infrastructure is provided by Vultr.
📺 Source: DeepLearningAI · Published May 22, 2026
🏷️ Format: Workflow Case Study
