First, create a model that is able to predict what character appears given an image. Use the model-shell.py file as a guide.
Once you get a sufficiently accurate model, create a program that takes in an image of a document that is able to go through and find all the letters on the page.
Good luck!