cosmo-wander-ai/cosmo-edge
Production-grade C++ edge AI engine for video analytics with visual pipeline orchestration, on-device VLM, and real-time OSD.
What it solves
CosmoEdge addresses the complexity of deploying AI models from files to real-world production environments. It solves the challenges of managing multi-channel video processing, hardware-accelerated inference, and real-time event reporting on resource-constrained edge devices.
How it works
The system uses a C++17 native runtime to handle high-performance tasks like video decoding, inference scheduling, and OSD rendering. Users can build complex workflows using a browser-based visual pipeline editor that connects video sources, AI models, and alarm rules. The engine supports various architectures, including x86 for development and Sophon NPUs for production, and can export structured data via MQTT or HTTP webhooks.
Who it’s for
It is designed for developers and deployment teams building industrial-grade video analytics applications, such as safety monitoring, automated visual inspection, and object counting systems.
Highlights
- Visual Orchestration: Build and manage AI workflows through a web-based visual pipeline editor.
- Hybrid VLM Support: Supports prompt-driven vision models like GroundingDINO and VLMs (e.g., Qwen series) via on-device or API-based backends.
- Production-Ready Runtime: C++ native core optimized for low-overhead, continuous multi-channel video workloads.
- Real-time Debugging: Includes an OSD system for business labels, semantic color coding, and zone overlays.
- Flexible Integration: Supports RTSP, USB cameras, and WebRTC, with data export via MQTT and HTTP.