Built an end-to-end pipeline to classify cognitive load using instruction-free eye-tracking data. Engineered 150+ gaze features (fixations, saccades, blinks, entropy, pupil metrics). Benchmarked ML models, achieving 93% accuracy with Random Forest. Developed a Hybrid TCN + ViT deep learning model, achieving 98.8% accuracy & 0.99 F1-score. Applied SHAP and Grad-CAM++ to interpret feature importance and temporal attention. Validated generalisation using subject-wise splits across 52 participants (196 + 80 trials).