In my first year at IIT Bombay, I wanted to learn machine learning by building something real. Through IITB’s Seasons of Code program (where seniors mentor students end-to-end), I worked on a real-time sign language recognizer.
The system runs in a live pipeline: webcam feed → hand landmark extraction (MediaPipe) → normalized feature vector → TensorFlow model → predicted sign. I focused heavily on preprocessing so the model stays robust across different hand sizes, positions, and camera distances, and added temporal smoothing to reduce flicker in live predictions.
The final prototype achieved ~97% accuracy on the evaluation split and worked reliably in real-time on a laptop webcam, even under noisy frames like partial hands or fast motion. Building this taught me core ML fundamentals early—data pipelines, training/evaluation loops, and shipping inference outside notebooks under real-time constraints.