Abstract : Driver fatigue constitutes a leading causal factor in fatal road traffic collisions worldwide, with drowsiness-impaired driving implicated in a substantial fraction of crashes occurring on monotonous highway segments during overnight and early-morning hours. This paper presents a multi-modal AI system for real-time drowsiness detection that fuses visual behavioral markers, eye aperture dynamics quantified via the Percentage of Eye Closure (Perclos) metric, head pose trajectory, and yawning frequency, with physiological signals comprising electroencephalographic (EEG) slow-wave activity and heart rate variability (HRV), and with vehicle dynamics inputs including lateral deviation and steering entropy. Convolutional neural networks process visual streams; lightweight edge-deployable classifiers integrate EEG and HRV features; a sensor fusion layer applies Dempster-Shafer evidential reasoning to yield a calibrated drowsiness probability estimate with sub-second latency. Detection performance achieves 94.7% accuracy and 0.96 area under the receiver operating characteristic curve across a held-out evaluation corpus. A three-level graduated safety intervention protocol, auditory-haptic alert, speed reduction recommendation, and autonomous vehicle handover initiation, is specified with intervention thresholds derived from drowsiness-progression modeling. Regulatory alignment with NHTSA Advanced Driver Assistance System guidelines and ISO 26262 functional safety standards is discussed.