DNN's make good features though, especially if you have time series data or lots of text.

I agree that the final model should be a randomforest/xgboost/lightgbm for typical tabular data.