Built a real-time gaze-controlled cursor at 30+ FPS using MediaPipe Face Mesh with head-pose compensation, reduced jitter with Kalman filtering + exponential smoothing, and designed a predictive virtual keyboard with dwell-click input.

Built an end-to-end ML pipeline on the Cleveland Heart Disease dataset and deployed Django-backed prediction serving with PostgreSQL records, secure form validation, and role-based data access.

Benchmarked 6 architectures (custom CNN, ResNet50, EfficientNetB0, VGG-16, and tuned variants). Best model (ResNet50 Tuned) achieved 98.77% accuracy and 98.76% F1-score across 4 tumor categories.

Automated 10+ OS-level tasks through voice commands using SpeechRecognition, Google Speech API, and custom intent parsing for accurate task execution.