As digital education evolves, keeping an eye on student engagement in online classes becomes challenging. This study proposes a system that will detect the attention and distractions of students in real-time by utilising multi-modal framework based on deep learning. This system is built using various pre-trained models. These are YOLOv8 for objects detection, MediaPipe for facial landmarks and pose estimation, and DeepFace for facial emotion recognition. Classification of attention states is accomplished through XGBoost using features that include the emotion, eye, and sleeping postures, object-based distractions, and face direction. Other logics like object-based distractions, facial orientation, face blurring and head pose analysis makes the system robust. Each distraction generates alert and logging in real time, while report at end of session summarizes the attention trend. Testing done with many videos and datasets show that the system detects distraction types and inattention very accurately. Thus, system can be used in virtual learning systems. This framework can be integrated into educational platforms for better engagement and tailor-made feedback.