Adaptive learning platforms generate extensive student data, enabling the prediction of academic risks, but often fail to provide students with actionable, personalized guidance for improvement. This paper introduces a novel framework that leverages machine learning for dropout prediction and Large Language Models (LLMs) for generating natural language, personalized learning recommendations. Our primary contribution is a highly accurate dropout prediction system that achieves an AUC-ROC of 0.908, accuracy of 87.4%, and precision of 0.981 using Random Forest classification. Using the “Adaptive Learning & Personalized Education Dataset,” we build predictive models to identify at-risk students, and these insights are then synthesized by an LLM to produce specific, coherent advice on intervention strategies. Since direct deployment with students was not feasible, we established a rigorous expert-based evaluation methodology to validate our system’s output. In this study, a panel of experienced educators assessed the LLM-generated recommendations across key dimensions of relevance, actionability, and pedagogical soundness. The results demonstrate that the experts consistently rated the AI-generated guidance as high-quality and contextually appropriate, confirming the viability of combining accurate dropout prediction with LLMs as a scalable tool for creating pedagogically valid support. This work not only presents a functional pipeline for generative educational advice but also provides a critical methodological blueprint for validating such AI-driven interventions in the absence of live student trials.