MyGrad

An autograd engine and neural network built from scratch, with interactive training visualization.

Website · GitHub

Built a complete automatic differentiation engine and neural network from scratch, inspired by Karpathy’s micrograd. Includes an interactive web visualization that lets users scrub through training steps and watch decision boundaries evolve in real time.

Features:

  • Custom Value class implementing backpropagation and autograd
  • MLP with configurable layers and ReLU activations
  • Real-time visualization of decision boundaries across 300 training epochs
  • Interactive web UI with loss and accuracy tracking charts
  • Achieves 100% accuracy on the make_moons binary classification task

Technologies: Python, JavaScript, HTML5