About the platform
GradVex turns a neural network into a visible, playable system.
The project is built as an educational full-stack-style web app for the AI era: a live playground, visual architecture explorer, math panel, 3D scene, and guided explanations that make weights, biases, activations, loss, and predictions concrete.
Learn by doing
Every concept connects back to a real action in the playground.
Local and fast
The trained model runs in-browser after the weight files load.
Inspectable math
Users can move from intuition to formulas without leaving the app.
Model details
Architecture
MLP 784 -> 128 -> 64 -> 10
Parameters
109,386
Training data
MNIST, 60k train / 10k test
Loss
Categorical cross-entropy
Inference
Runs locally in the browser
Purpose
Explain neural networks through interaction
Technology stack
Next.js 16App Router and production build
React 19Interactive client components
TensorFlow.js 4In-browser neural network inference
Three.js3D network visualization
ZustandShared UI and model state
Tailwind CSS 4Responsive design system