We investigate the integration of semantic analysis techniques and machine learning models, to identify and predict academic success patterns for high school students. Starting from an existing dataset, we generated summary notes for teachers based on key academic indicators, and transformed them into embeddings using a lightweight transformer model (DistilBERT). Principal component analysis was applied to reduce the dimensionality of the embeddings, to be then used for K-Means clustering. A decision tree classifier was trained to predict student success, leveraging both classical features (such as grades, non-attendance, and failures) and semantic embeddings. The results show the great potential of combining structured and unstructured data for early detection of students at-risk.