TSC walkthrough. Manage video streams on EdgeLake with local AI inference, follow similar results across multiple cameras, and configure models without sending video to the cloud.
For customers asking how to manage video at the edge without centralizing it, the AnyLog team built a TSC demo of EdgeLake handling live video streams from two cameras pushed into a single edge node. The node runs EdgeLake alongside a standard MongoDB container for object storage and a SQL database for the structured inference data.
The architecture is straightforward. When video lands on the edge node, EdgeLake hands each frame to a YOLO model running locally. (The team published a GitHub repo with a Docker GPU container so anyone can clone and run the same setup.) Inference results are written to the SQL database. The raw video file is written to the MongoDB object store. That split is what makes the data discoverable later: queries hit the metadata-rich SQL side and surface the matching video clips, instead of relying on object-storage prefix scans.
For scale, the demo pattern dedicates a separate AI node with a GPU sitting next to the EdgeLake operators. Cameras stream to the edge nodes, the edge nodes forward frames to the AI node, and inference comes back over the same hop. More cameras, more EdgeLake nodes, more AI nodes. The topology stays flat and SQL-addressable.
The same setup also handles multi-tenant data ownership. In the chemical-plant scenario discussed during Q&A, each company owns the nodes inside its production lines and decides what data is shared and what stays private. Infrastructure-level telemetry like shared electrical distribution can be exposed across companies; per-product production data stays scoped. All of it is queried over the same SQL interface, with per-view permissions per node.