Baked our own image models that process each frame to detect objects. Then we added some lightweight video models to detect actions and tracking