Personalizing STEM Learning at Scale: An Adaptive AI Tutoring Suite with Student Modeling, Curriculum Sequencing, and Explainable Formative Feedback

Personalized, real-time instruction in STEM classrooms remains a persistent challenge: student cohorts vary widely in prior knowledge and misconception patterns, while teachers lack timely, individualized data to adapt instruction dynamically. This paper presents the Adaptive Tutoring Suite (ATS), a modular AI-driven platform designed to improve measurable learning outcomes in STEM micro-learning environments. The ATS integrates four evidence-informed components: (1) a hybrid Bayesian Knowledge Tracing (BKT) and transformer-embedding student model that produces calibrated, real-time mastery estimates; (2) a constrained curriculum sequencing engine combining Mixed-Integer Quadratic Programming (MIQP) with an online contextual bandit to minimize expected lessons-to-mastery under classroom resource constraints; (3) an explainable automated formative feedback pipeline that fuses large language model (LLM) outputs with symbolic validators to generate accurate, actionable hints; and (4) a teacher-facing dashboard providing intervention recommendations and a feedback review queue. We evaluate the ATS using publicly available datasets (ASSISTments, EdNet) and parameterized synthetic classroom cohorts, reporting mastery prediction accuracy, normalized learning gain, time-to-mastery reduction, and feedback utility. Results from simulation and ablation analyses indicate that the hybrid student model improves mastery calibration over standard BKT, the MIQP-bandit sequencer reduces expected instructional time compared to greedy baselines, and the hybrid feedback pipeline achieves higher teacher-rated hint utility with fewer high-confidence errors than an LLM-only approach. Implications for scalable, equitable personalization in STEM education and directions for real-world classroom deployment are discussed.