← Back to Projects
In-Browser Neural Network (From Scratch)
A digit-recognition network hand-built in raw NumPy, int8-quantized to 145 KB, running as pure JavaScript. Try it live.
NumPyNeural NetworksQuantizationJavaScriptReact
Overview
A complete neural network built without any ML framework: hand-written forward and backward passes, Adam optimizer, and shift augmentation in raw NumPy, reaching 98.2% test accuracy. The weights are int8-quantized (4× smaller, no accuracy loss) and shipped as a 145 KB JSON file, with inference running as ~80 lines of dependency-free JavaScript directly in the visitor's browser. A parity test verifies the JavaScript engine reproduces the Python model's probabilities to within 1e-6.
Key Features
- ✓Trained from scratch, no TensorFlow, no PyTorch
- ✓98.2% test accuracy, preserved after int8 quantization
- ✓Zero-dependency JavaScript inference in the browser
- ✓Python↔JS parity verified to 1e-6