Real-time ASL word translator that runs entirely in the browser. Uses MediaPipe Hands for landmark tracking and a stacked LSTM model converted to TensorFlow.js for fully client-side inference.
- Engineered a real-time American Sign Language (ASL) translator capable of classifying 40 distinct words with a 67.5% top-3 test accuracy by training a stacked LSTM sequence model on the WLASL dataset
- Eliminated backend server dependencies and ensured user video privacy by converting the Keras LSTM model into TensorFlow.js for fully client-side browser inference
- Developed a reactive frontend interface to process live webcam feeds, extract 3D hand landmarks in real-time, and stream data directly into the browser-based inference engine