Looks really amazing! Reproducibility of models and experiments is huge. It should be almost a requirement if one is going to publish results claiming SotA, etc. Seems like you could become the GitHub for deep learning in addition to the Heroku for deep learning.