Deep-learning courses hand you MNIST on a plate. Persian digits — ۰ through ۹ — had no such plate. So the final project for the Deep Learning course at Shahid Beheshti became two projects: build the dataset, then learn from it.
The dataset — handwritten Persian digits, collected and preprocessed into an MNIST-style corpus.
The model — a neural network trained and evaluated on that corpus, in Python.
The lasting lesson was the first half: real machine learning starts long before the model, in the data nobody has gathered yet.