What was built
- Live hand tracking with MediaPipe Hands; gestures classified by a MobileNetV2-based model.
- Recognises 36 classes (A–Z, 0–9) from a webcam feed and converts them to text with confidence scores.
- Confidence thresholding and smoothing for stable predictions; landmarks and bounding boxes drawn on the video.
- Improved validation accuracy from 72.8% to 85.2% through augmentation, feature extraction and hyperparameter tuning.