We're using a pretrained model on tensorflow. However, we're planning to make it in-house in the near future.