Portfolio Project
Shape Classifier Demo
Handwritten Shape Recognition
Context
I wanted to create a model that recognizes handwritten shapes.
Approach
- Downloaded images from Google's QuickDraw dataset to build training and validation splits.
- Trained a compact ResNet18 using PyTorch Lightning.
- Deployed a minimal AWS Lambda handler for serverless CPU inference from the browser.
Impact
- Predicts circle, triangle, square, hexagon, or octagon from a single drawing with about 90% accuracy.
- Demo shows responses return in under a second after a 10‑second warm‑up.
