Interactive inference demo

Draw a digit and inspect what the network sees

The canvas is now the default starting point. Sketch a number, run inference, and compare your raw input with the normalized 28x28 image sent into the model.

Sketch a single digit

Draw directly on the canvas, keep the number centered, and use the prediction button to send it through the MNIST preprocessing and classification pipeline.

Single digit Centered strokes High contrast

Upload an existing image

If you already have a handwritten digit image, upload it here and the same preprocessing pipeline will convert it before inference.

Drag and drop an image here

or browse files

Supported formats: PNG, JPG, JPEG

How the project works

From browser strokes to a predicted digit

This project is a compact inference demo built around an MNIST-style handwritten digit classifier. What you draw is transformed into the same representation used during training.

Capture

The browser reads your drawing from the canvas or takes an uploaded image as input.

Preprocess

The backend converts the image to grayscale, resizes it to 28x28, normalizes pixel values, and inverts if needed.

Infer

The trained neural network consumes the flattened 784-value input and computes class logits for digits 0 to 9.

Explain

The UI shows both the processed image and the full probability chart so you can see why one digit won over the others.

What is happening mathematically?

Each dense layer learns useful stroke patterns such as loops, edges, angles, and relative placement. The last layer produces scores, and softmax turns those scores into probabilities.

Best input tips

Draw one digit at a time.
Keep the digit close to the center of the canvas.
Use thick, clear strokes and avoid extra marks.

Draw a digit and inspect what the network sees

Sketch a single digit

Upload an existing image

Prediction Results

Original Input

Preprocessed (28×28)

Prediction

Probability Distribution