Real-time computer vision at a sports venue looks simple on a whiteboard: cameras produce video, models detect players and the ball, software reacts by steering a camera, cutting between angles or generating stats. The first design most teams draw sends every feed to the cloud, because that is where the GPUs and the MLOps tooling already live. That design fails on physics and economics before it fails on code. This post walks through how we split such a system between an on-site edge and the clo...