Projects
← Back to all projects2026 / Live
MNIST Digit Recognition.
Library-free neural network in pure TypeScript and JavaScript that classifies handwritten digits from MNIST. Backpropagation and forward propagation implemented from scratch.
- Role
- Author
- Status
- Live
- Year
- 2026
A full digit-recognition neural network without TensorFlow, PyTorch, or any other ML library. Forward and back propagation are implemented from scratch in TypeScript and JavaScript, including activations, loss, and gradient updates.
The training backend runs on Node.js. The web frontend offers an HTML5 canvas, and drawn digits go through a MNIST-style centre-of-mass normalisation and a 28 by 28 pixel resampling step before being fed to the network.
On a 50,000 train and 9,000 test split the network reaches roughly 95 percent precision. A learning project whose value lies precisely in not using a framework.