We tested two model architectures:

vanilla neural network with Dense layers and feature extraction convolutional neural network with tokenization of each individual letter, padding, max-pooling

The Accuracies we got was:

Vanilla NN: ~71.9% CNN: 96.2%