Point the camera at a person, cat or dog. A pose model finds their joints and the GPU draws a skeleton that follows them, over a radiograph version of the camera feed. Everything runs on this device; no video leaves it.
Human pose: MediaPipe Pose Landmarker (lite), with Face Landmarker for the head and jaw. Animals: EfficientDet-Lite0 finds the cat or dog, RTMPose-m (AP-10K) finds 17 joints. Rendering: WebGPU. The skeleton is a drawn template fitted to the joints, not a scan.