Ask HN: Is Lightning the PyTorch equivalent of Keras for TensorFlow?

I’ve been recently approached by a small, innovative, industrial AI startup still in search of their best pipeline, in relation to TensorFlow vs PyTorch… or, better, Keras vs Lightning. There are a number of tutorials and blog articles explaining the differences, but we are here talking about the best bet for the next 1-3 years. This seems a good question for HNers then: what do you think about? For what I see, Lightning is really poised to become for PyTorch what Keras is for TensorFlow, that allowing that small, industrial AI startup to focus in full on their domain cases while using both high-level wrappers and both the underlying libraries. Exciting times ahead? Thanks.