Real-Time Sign Language Translator

The problem

A live alphabet recognizer has to connect camera tracking, model input, inference, and readable feedback. I built this project to take a TensorFlow classifier beyond a training script and make its predictions usable in a browser.

What I built

• The camera and translation interface. React displays the webcam feed and tracked hand landmarks, sends prediction requests, and presents the returned letter and confidence in a translation console.
• A landmark-based training pipeline. MediaPipe supplies 21 hand points, each with x, y, and z coordinates. Those 63 features form the input to the TensorFlow classifier and its converted TensorFlow Lite model.
• A focused inference API. Flask serves the lightweight model and returns a prediction with its confidence score. Training and conversion remain separate from the deployed inference runtime.

Why I built it this way

• Use hand geometry as the model input. Landmarks give the browser and API an explicit feature contract: 21 points become one 63-value vector per prediction.
• Keep inference separate from training. A converted TensorFlow Lite model serves predictions while the extraction and training scripts remain available for improving the dataset and classifier.
• Show confidence and withhold weak predictions. The interface gates low-confidence results rather than always displaying a letter. Similar hand shapes still need better data and temporal modeling.

What I learned

I practiced the complete ML application path: extracting features, training and converting a model, serving inference, and connecting it to a live interface. Consistent input features and clear feedback matter as much as a working prediction endpoint.