Inference runs entirely on this device via WebAssembly. The server only delivers the static app shell and (once) the model file — it does not see the video.