Instacart Kaggle competition just got finished and 3rd place winner was pure deep learning approach with no hand made feature engineering.The problem was really interesting and is written deep learning all over it, but many people including myself ended up using gbms with hand made features.Here is the 3rd place winner post
I was just going to post this myself! This is a fascinating solution and I am definitely going to be looking at it in detail.
Fascinating indeed, he used a combination of wavenet + convolutinal-LSTM. I think he is the only guy who did pure deep learning without any hand made features. Everyone else including top 10 did traditional feature engineering
Here is another forum post outlining a solution based on RNNs.