I'd like to turn this into a product, so I don't want to release the code. I wrote it in Theano (Python).

Not the author, but I'm guessing it uses an implementation of "Grammar as a Foreign Language"[1].

TensorFlow has a tutorial that gets you most of the way[2].

Edit: playing a bit I don't think it is doing this. I think it's using Word2Vec style vectors to build similar sentences, and calculating the probability of each and maybe the edit-cost to change.

[1] http://arxiv.org/abs/1412.7449

[2] https://www.tensorflow.org/versions/master/tutorials/seq2seq...