GT-AiVision Deployment Architecture
Edge-Native Real-Time Detection With Flexible VLM Deployment
Connect compatible RTSP cameras to NVIDIA Jetson edge processing, the GT-AiVision platform, event and incident data, dashboards, alerts, and a configurable local or remote Vision Language Model.
Architecture overview
Qualify each layer before committing to scope
The final design depends on the customer camera environment, network, required real-time channels, Virtual Patrol coverage, VLM choice, retention, integrations, and support model.
IP Camera / NVR / VMS
Stable, compatible RTSP streams on the customer network.
NVIDIA Jetson
Decode, inference, tracking, ROI, tripwire, event rules, and snapshots.
FastAPI + Data
Cameras, events, patrols, users, reports, settings, and incidents.
Local or Remote VLM
Virtual Patrol analysis and optional natural-language event descriptions.
Operators and Systems
Dashboards, alerts, reports, incidents, APIs, and support workflows.
Camera layer
Compatible IP cameras or systems provide stable RTSP streams. Continuous video remains within the customer CCTV/NVR environment unless another arrangement is explicitly designed.
Jetson edge layer
Independent DeepStream workers use NVIDIA hardware decoding, YOLOv11/TensorRT inference, tracking, ROI and tripwire analytics, event rules, snapshot capture, and event recording.
Platform and data layer
The FastAPI backend provides camera, event, Virtual Patrol, reporting, user, settings, system, and Incident Management functions. MySQL stores structured platform records.
VLM layer
Virtual Patrol and optional event descriptions use a configurable local or remote VLM selected according to privacy, performance, hardware, commercial, and operational requirements.
Operator layer
Authorised users access dashboards, status, events, snapshots, patrol findings, reports, incidents, settings, and system controls through the web interface.
Integration layer
API-token support, Telegram, SMTP, and customer-specific integrations are applied according to the qualified network and support design.
Typical data flow
Real-time detection and Virtual Patrol use different processing paths
Connect
A detection worker connects to the configured RTSP stream.
Process
DeepStream decodes video, runs inference, tracks objects, and evaluates rules.
Record
A triggered rule saves a snapshot and structured event record.
Describe
An optional VLM worker can attach a natural-language scene description.
Notify
Telegram can send the configured event and snapshot.
Review
Operators review evidence through the dashboard.
Escalate
Relevant events can become managed incidents.
Patrol separately
Virtual Patrol captures scheduled snapshots, uses the configured VLM, records findings, and generates reports.
Connectivity model
Core detection can be local. The complete service design may still use external connectivity.
Can run on the customer LAN
- Real-time object detection and tracking.
- ROI and tripwire evaluation.
- Event-rule processing and snapshot capture.
- Local database and dashboard functions, depending on deployment.
- Local VLM where a qualified design is selected.
Requires external connectivity when used
- Telegram alerts.
- SMTP email delivery.
- Remote VLM services.
- Remote technical support.
- External API integrations outside the customer LAN.
Do not describe the complete platform as having no cloud or internet dependency. The answer depends on the selected functions.
Camera and network requirements
RTSP compatibility must be verified
“RTSP-compatible” does not mean every camera, recorder, codec, authentication method, or network configuration works without assessment.
Qualification checklist
- Stable RTSP stream from the camera, NVR, or VMS interface.
- Reliable LAN and fixed or reliably assigned camera addresses.
- Supported codec and suitable resolution.
- Camera angle, lighting, object size, and scene visibility.
- Power and physical location for the Jetson appliance.
- Customer authorisation for camera and image processing.
- Agreed firewall, remote-access, retention, and support arrangements.
Package architecture
Separate managed-camera coverage from real-time processing
Managed-camera coverage supports platform configuration and Virtual Patrol design. It is not simultaneous real-time inference across every managed camera.
| Package | Hardware | Managed-camera coverage | Maximum listed real-time channels |
|---|---|---|---|
| Aivision-Nano | Jetson Orin Nano | Up to 100 cameras | Up to 8 |
| Aivision-NX | Jetson Orin NX | Up to 300 cameras | Up to 16 |
| Aivision-AGX | Jetson AGX Orin | Up to 500 cameras | Up to 24 |
These are package maxima, not unconditional performance guarantees. Actual suitability depends on model, resolution, frame interval, tracking, rules, VLM use, stream stability, and scene complexity. Final sizing requires technical qualification.
Platform technologies
A practical edge-AI and web-platform stack
Edge AI
NVIDIA Jetson Orin, DeepStream 7.1, YOLOv11, TensorRT, hardware decoding, and tracking.
Application
Python, FastAPI, Jinja2, HTML, JavaScript, and Tabler UI.
Data and services
MySQL databases and systemd services for per-camera workers.
Notifications and VLM
Telegram Bot API, SMTP email, and configurable local or remote VLM service.
Technology names describe the implementation and do not imply a formal vendor partnership unless separately approved.
Documented platform controls
Application controls support the deployment security design
GT-AiVision includes authentication, permission, service-health, retention, and update functions. These do not replace project-level network and operational controls.
- JWT authentication and roles.
- API-token support and user administration.
- Email password reset.
- Configurable snapshot and log retention.
- GPU, memory, worker, and service-health monitoring.
- Per-camera service controls.
- Software version tracking and update mechanism.
Deployment security recommendations
Define infrastructure and compliance controls per project
These are implementation recommendations, not cybersecurity certifications. Customers and partners remain responsible for their regulatory, privacy, and organisational requirements.
Singapore deployments should include a PDPA review; regional deployments require consideration of applicable local rules.
Recommended project controls
- HTTPS termination and approved encrypted remote access.
- Firewall rules and network segmentation.
- Restricted camera and database access.
- Strong account and permission administration.
- Customer-approved retention and backup policy.
- Site notices, privacy review, and authorisation.
- Logged support and administrative access.
Partner deployment responsibilities
Agree the responsibility matrix before installation
Partner-led activities
- Customer discovery and site coordination.
- Camera, NVR, and VMS information collection.
- Network, power, and physical installation coordination.
- Local installation assistance where agreed.
- User coordination and first-line support.
- Local privacy, language, and operational requirements.
Grep Tech activities
- Product and deployment training.
- Technical qualification guidance.
- Platform setup and configuration support.
- Software updates and maintenance.
- L2/L3 investigation and escalation.
- Integration and customisation scoping.
The final responsibility matrix is confirmed in the partner agreement or project quotation.
Deployment process
Move from discovery to supported operation
Discover
Confirm cameras, network, use cases, alerts, users, reporting, and integrations.
Qualify
Review representative streams, views, lighting, zones, and event definitions.
Design
Select hardware, patrol coverage, real-time channels, VLM, retention, and support.
Propose
Confirm hardware, subscriptions, installation, logistics, taxes, integrations, and responsibilities.
Deploy
Install the appliance and configure cameras, rules, users, notifications, and reports.
Test
Validate representative scenarios and tune supported configuration parameters.
Handover
Train users and document support and escalation contacts.
Support
Apply updates and escalate issues through the agreed process.
No fixed installation time is promised until the customer site and project scope are qualified.
Integration options
Scope APIs and workflow changes around real project requirements
GT-AiVision provides a FastAPI backend and API-token support. Integration possibilities depend on required data, workflow, authentication, network access, third-party systems, and support responsibility.
Technical review is required for
- Custom API work.
- Customer-specific reports.
- Notification integrations.
- Workflow changes.
- External authentication and network access.
Additional work may be quoted separately.
Technical partner discussion
Qualify the customer environment before committing to scope
Review cameras, network, coverage, real-time channels, VLM options, integration, security controls, and delivery responsibilities.