keras-team/keras ? reverse-engineered prompt

Build me a clean, production ready version of Keras 3, a deep learning library that feels simple for humans to use but can run on top of JAX, TensorFlow, PyTorch, and OpenVINO for inference.

I want the core library to make it easy to define models, train them, save and load them, and run common workflows for things like computer vision, text, audio, forecasting, and recommender systems. Keep the experience friendly for people coming from tf.keras, and make sure the backend can be chosen before import through an environment setting or config file.

Please include good examples and clear docs so someone can get started quickly, and make sure the install and local dev flow works well from the README. If you need to check current backend docs or APIs while wiring things up, look them up online. Also make sure the codebase is organized for tests, benchmarks, and backend compatibility, since I want this to work reliably across the supported frameworks.