Facial Emotion Recognition (FER) and engagement analysis are critical for applications in human–computer interaction and health care. This study presents a robust model leveraging Vision Transformers (ViTs) for emotion classification and K-Means clustering for engagement level categorization. ViTs process images as sequences of patches, capturing both global and local dependencies for superior feature extraction, while K-Means clustering classifies engagement levels into three categories: highly engaged, moderately engaged, and barely engaged. The unsupervised nature of K-Means enables dynamic adaptation to varying user interactions without requiring manually labeled data. Experimental results on the AffectNet dataset demonstrate an 84% accuracy in emotion recognition, with engagement classification validated by silhouette scores exceeding 0.78. This approach introduces a novel methodology to classify emotional engagement, advancing FER systems and providing valuable insights into real-time human behavior analysis.