vietanhdev/tf-blazepose ? reverse-engineered prompt

Build me a TensorFlow 2 app for real time human pose detection based on BlazePose. I want it to be able to train on pose datasets, run on a webcam or video file, and show the detected body keypoints on top of the person in the frame. Please make it feel complete enough to use for a simple workout counter too, so the pose results can be read clearly and reused later for counting movements like pushups.

Use the existing idea of separate configs for different training setups, and make sure there is a straightforward way to train, test, and run the demo from the command line. If you need to, look up the current docs for BlazePose or TensorFlow 2 model loading so the implementation matches modern usage. Keep it focused on the pushup style keypoints if that is the easiest supported dataset path, but structure it so other pose datasets could be added later.