This document summarizes an optical character recognition (OCR) system that uses unsupervised feature learning and sequence-to-sequence modeling. It proposes using stacked restricted Boltzmann machines to learn features from unlabeled text image data, and encoding the text images into fixed-dimensional representations using a recurrent encoder. A recurrent decoder is then used to map these representations to text predictions. The system is evaluated on five Indic scripts, achieving error rate reductions over traditional profile-based features. Future work directions include generative modeling of representations and attention-based sequence modeling for OCR.