This app uses a convolutional neural network (CNN) to recognize handwritten digits and letters. The neural network was trained using the EMNIST dataset (EMNIST ByClass) and achieved a test accuracy of ~87%.
Unfortunately, the dataset is quite unbalanced, so the model tends to be biased.
To minimize inference errors, draw the digit or letter in the most generic form possible: upright and with the right proportions for the character to fit inside a square box. Capital letters tend to be more accurate than lowercase letters.