Keep inference on the edge
Video decode, preprocessing, and multiple computer-vision pipelines run locally on Jetson, limiting dependence on centralized raw-video processing.
Production AI and Private Infrastructure
SentinelAI - Secure Edge Video Intelligence Platform
A secure Jetson-based edge video platform combining multi-stream computer vision, custom Yocto Linux, mTLS device management, WebRTC, and post-quantum-protected media content.
Edge AIEmbedded LinuxVideo systemsComputer visionSecure communications
The system
SentinelAI is an edge-computing platform built around NVIDIA Jetson NX 16GB devices. Each device can ingest multiple local USB cameras and remote WebRTC streams, process video locally through GStreamer and NVIDIA DeepStream, and run multiple computer-vision models concurrently without sending every source stream to central infrastructure.
Representative capabilities include face detection and recognition, license-plate detection and recognition, document detection and recognition, and OCR. YOLO26, Tesseract, and additional specialized models contribute to the processing path; selected models, recognition logic, thresholds, and identity-related implementation details are intentionally not disclosed.
The device software runs on a custom Yocto-based Linux platform and maintains an authenticated management relationship with a central system. Real-time remote media uses WebRTC, with an additional post-quantum protection layer inside the media-content path and a modified SFU capable of handling the protected flow without requiring ordinary access to the underlying unprotected video content.
Engineering relationship
Nine OPTIME engineers contributed across a one-year program spanning embedded Linux, media pipelines, GPU inference, computer vision, WebRTC, SFU adaptation, device management, and protected-content integration.
OPTIME’s responsibility crossed the edge device and central communication boundaries: the custom operating-system platform, local multi-stream processing, model integration, machine authentication, protected remote-media path, and modified intermediary behavior had to operate as one system.
The current public case describes the completed engineering scope conservatively and does not claim a specific downstream deployment scale or certification.
Engineering constraints
- Run multiple concurrent camera and model pipelines within the compute, memory, thermal, and I/O limits of NVIDIA Jetson NX 16GB.
- Combine local USB-camera capture and remote WebRTC input within one edge media architecture.
- Maintain local inference for sensitive video and identity-related signals rather than depending on continuous central processing.
- Operate custom Yocto Linux, GStreamer, DeepStream, GPU inference, and device-management software as one maintainable appliance stack.
- Authenticate edge devices to the management platform through mTLS without making unsupported compliance claims.
- Protect video content inside the WebRTC media path while allowing a modified SFU to route the protected format.
What OPTIME engineered
- A custom Yocto-based Linux platform for the Jetson edge appliance.
- GStreamer and NVIDIA DeepStream pipelines for multiple USB-camera and WebRTC video inputs.
- Local decode, preprocessing, and concurrent computer-vision inference on the NVIDIA GPU stack.
- YOLO26, Tesseract, and specialized-model integration for face, plate, document, and OCR workflows.
- Event and metadata generation from local video-intelligence pipelines.
- mTLS-authenticated machine-to-machine communication with the central management platform.
- A post-quantum protection layer inside the WebRTC media-content path.
- SFU modifications enabling intermediary handling and routing of the protected media format without ordinary access to unprotected video content.
Architecture
USB cameras and remote WebRTC streams
Multiple local and remote video sources enter the edge device.
GStreamer media pipelines
Capture, transport, decoding, and pipeline control unify the different video inputs.
NVIDIA DeepStream preprocessing
GPU-aware decode, preprocessing, batching, and model integration prepare streams for local analysis.
Concurrent CV inference
YOLO26, Tesseract, and specialized models support face, plate, document, and OCR workflows.
Metadata and events
The edge device emits bounded recognition and detection outputs rather than exporting every raw processing detail.
Protected WebRTC content path
A post-quantum protection layer operates inside the real-time media-content flow.
Modified SFU
The intermediary handles the protected media format without requiring ordinary exposure of the underlying video content.
Authorized remote endpoint
Protected video reaches an authorized destination through the real-time communications path.
mTLS management plane
A separate authenticated channel connects the edge appliance to central device management.
Key engineering decisions
Own the embedded platform
A custom Yocto Linux environment gives the appliance a controlled software foundation for media, GPU, model, update, and management responsibilities.
Separate media protection from WebRTC itself
WebRTC remains the real-time transport. OPTIME added post-quantum protection inside the media-content path rather than making an imprecise claim that the WebRTC protocol was replaced.
Adapt the intermediary to protected media
The SFU was modified to understand and route the protected format without requiring ordinary access to the unprotected video content.
Platform evolution
The engineering program brought embedded software, multi-stream media, local computer vision, secure management, and protected remote video into one edge platform.
Specific customer topology, security implementation, model configuration, recognition thresholds, and deployment details remain intentionally generalized.
Verified capability
The platform supports local processing of multiple USB-camera and WebRTC streams on NVIDIA Jetson NX 16GB, with representative face, license-plate, document, and OCR workflows operating through several concurrent models.
Central management uses mTLS-authenticated machine communication. Remote video uses WebRTC with post-quantum protection inside the media-content path and a modified SFU designed to handle the protected flow. No claim is made about compliance certification, universal security, or downstream deployment scale.
Technology & Engineering Role
- NVIDIA Jetson NX 16GB
- Edge compute target for multi-stream video processing and local inference.
- Yocto Linux
- Custom embedded operating-system foundation for the managed appliance.
- GStreamer
- USB-camera and WebRTC ingest, decoding, and real-time media pipelines.
- NVIDIA DeepStream
- GPU preprocessing and integration of concurrent video-inference pipelines.
- CUDA / NVIDIA GPU stack
- Local accelerated execution for supported computer-vision workloads.
- YOLO26
- Object-detection foundation within selected visual workflows.
- Tesseract
- OCR component for document and text-recognition workflows.
- Specialized CV models
- Additional face, plate, and document processing kept general for confidentiality.
- WebRTC
- Real-time transport for remote video sources and authorized endpoints.
- Modified SFU
- Routing of the protected media format through intermediary infrastructure.
- Post-quantum media-content protection
- Additional protection inside the WebRTC content path; implementation details are withheld.
- mTLS
- Machine authentication between edge devices and central management.