Secondary-level biology students preparing for competitive examinations require timely misconception diagnosis, adaptive practice, motivational support, and feedback that remains grounded in the syllabus. This paper presents an AI-powered coaching and incentive framework for Sri Lankan Grade 13 English-medium Biology students studying the Microbiology unit. The central contribution is a five-layer architecture that connects student activity evidence, a 27-submetric learner profile, a hybrid intervention decision system, and a guided large language model coaching layer. The intervention system uses an expert rule gate to enforce eligibility, threshold, sequencing, and fatigue constraints, while a contextual-bandit ranker adapts intervention priorities when outcome data become available. The large language model is therefore used as a bounded coaching delivery layer rather than as an unrestricted intervention selector. Its tool-calling module can retrieve student progress, search unit materials, recommend quizzes, manage study plans, inspect badge progress, and record topic-confidence checks. The evaluation retained 132 complete paired records from a 167-student registered sample frame. Readiness increased from 52.09 to 58.90 on a 0–100 scale, with a mean gain of 6.81 points, t(131) = 6.27, p < .001, and dz = 0.55. The framework illustrates how expert knowledge, adaptive machine learning, measurable backend evidence, and LLM-based coaching can be combined without exposing students to an unguided chatbot.

AI in education,contextual bandits,Large Language Models,intelligent tutoring systems,biology education,adaptive interventions,learning analytics,incentives