Note for Mobile Devices: When operating this application on mobile, a separate, dedicated compute server must be built and configured in order to successfully run the AI model.
Autonomous zero-latency threat triage converting standard optical & thermal feeds into actionable intelligence.
Bridging the gap between legacy security infrastructure and next-generation artificial intelligence.
The core purpose of the E2C platform is to transform existing "dumb" optical and thermal surveillance networks into proactive, intelligent defense systems.
Currently, security relies heavily on human operators monitoring multiple screens, leading to fatigue and delayed responses. Our goal is to provide an automated, zero-latency threat triage system that uses state-of-the-art Vision-Language Models to detect anomalies instantly, requiring zero hardware upgrades to the edge cameras themselves.
Watch the system ingest live feeds, apply zero-shot VLM object detection, and trigger spatial alerts in real time.
Live stream ingestion, dynamic zero-shot detection, and drawing real-time geofences.
Under-the-hood look at the LocateAnything-3B inference pipeline executing Parallel Box Decoding.
Hardware-Agnostic C2 Dashboard
AI Threat Evaluation Overlays
A fully decoupled, modular system pipeline spanning edge video ingestion to browser-based C2 tactical views.
Ingests feeds from Legacy IP cameras, Thermal/IR Optronics, PTZ Masts, or Drones. Frames are buffered asynchronously via OpenCV to prevent blocking.
NVIDIA LocateAnything-3B performs zero-shot detection. Kalman filters apply persistent tracking IDs, and spatial rules calculate point-in-polygon overlaps.
React C2 dashboard receives WebSocket broadcasts with base64 video frames + threat metadata. Users draw interactive geofence polygons directly on the canvas.
Powered by NVIDIA LocateAnything-3B, dynamically detect unseen threats—like "armed individual" or "unattended baggage"—without requiring expensive dataset retraining.
Shattered the autoregressive bottleneck. Extracts dozens of dense bounding boxes simultaneously, delivering critical real-time processing at sub-100ms latency.
Draw interactive exclusion zones directly in the browser. Calculates continuous ray-casting algorithms for point-in-polygon checks, triggering instantaneous WebSocket alerts.
A highly optimized ByteTrack-style association model using Constant Velocity matrices + IoU. Maintains unique entity IDs across visual occlusions and crowds.
Seamlessly taps into existing infrastructure—analog cameras, RTSP streams, ONVIF IP systems—infusing legacy security networks with state-of-the-art AI.
Every detected threat is immortalized. Extracts high-resolution snapshots, calculates hashes for tamper evidence, and logs rich JSON metadata for audit trails.
FPS limiters have been completely removed for maximum throughput.
| Profile | Latency | Resolution | VRAM |
|---|---|---|---|
| Fast | ~14-18 ms | 640×640 | 7.8 GB |
| Hybrid | ~22-32 ms | 1024×1024 | 11.7 GB |
| Slow | ~60-85 ms | Native 4K | 18.2 GB |
Getting started with the E2C platform is simple. Clone the repository, download the VLM weights, and launch the servers.
NVIDIA GPU (RTX 3080+)