Driver drowsiness is a major contributor to traffic accidents due to degraded attention and prolonged reaction time. This paper presents an edge-based in-cabin monitoring system that combines a camera with a lightweight facial-landmark model to detect drowsiness in real time. Using the estimated landmarks, three geometric cues are computed: Eye Aspect Ratio (EAR), Mouth Aspect Ratio (MAR), and head pose. These cues are temporally smoothed and fused into a drowsiness score using a dual-threshold (hysteresis) mechanism to reduce false alarms during continuous operation. The system is deployed on an embedded AI platform and evaluated under multiple behavioral scenarios (alert driving, simulated drowsiness with prolonged eye closure, yawning, and talking) and under varying lighting conditions. Experimental results indicate an average processing speed of about 15 frames per second (fps) (worst-case latency \({<90}\)  ms) during long continuous runs, supporting practical integration into existing vehicles.