We use a neural network variational autoencoder architecture in order to encode font styles into a latent representation. With the help of this architecture we enable our machine to learn similiarities between font styles. This, in turn, is used to build our recommender system.