I think he meant for character recognition -- it's explained in the post that they went for an in-house dataset (from sudoku magazines I understand) instead of

I think he meant for character recognition -- it's explained in the post that they went for an in-house dataset (from sudoku magazines I understand) instead of MNIST. They ran into some issues, found the way to solve them, and improved their training set. This allowed them to reach 98.6% accuracy, and after a few updates to the app over 99%.