Draw a digit.
Watch AI recognize it.
A real trained neural network. Ten possible digits. One tiny experiment in machine learning.
Your input
Draw any digit from 0 to 9 with your mouse or finger.
The drawing is processed locally into 8×8 grayscale pixels and evaluated by the trained model.
Model prediction
See the network’s real output probabilities.
Waiting for a digit
Draw on the canvas or choose a sample, then press Recognize digit.
Small network.
Real machine learning.
Built with Python and PyTorch, then exported for live predictions. The same model architecture used in the training code performs the inference here.
8×8 grayscale
64 → 64 → 32
10 digit classes
Trained on the small scikit-learn Digits dataset. The 96.94% accuracy was measured on 360 held-out dataset images, not live user drawings. Freehand recognition is an experimental extension and may be less accurate because the handwriting style, size and preprocessing differ.
View the training code and methodology ↗