Differentiation at Scale: A Curriculum-Aligned, Offline-First Retrieval System for Overcrowded Classrooms in Rwanda

Secondary classrooms in Rwanda and much of Sub-Saharan Africa routinely enrol more students than a single teacher can differentiate for by hand, while Rwanda's Competence-Based Curriculum (CBC) is explicitly formative-assessment-driven and learner-centred. Existing AI-assisted differentiation tools close this gap for well-resourced, connected, English-medium classrooms, but are largely unusable where connectivity is unreliable, compute is scarce, and materials must track a specific national curriculum. We present a teacher-facing, offline-first differentiation engine that converts routine formative-assessment results into REB-curriculum-tagged worksheet recommendations across three proficiency tiers, using lightweight retrieval rather than generative inference so it can run on low-power, disconnected classroom devices. Since a live classroom deployment was not feasible on this paper's timeline, we validate the system's core mechanics through a simulation-based computational pilot: a synthetic class of 68 students, 16 REB-aligned competencies, and a 48-item worksheet bank, evaluated under realistic formative-assessment noise and incomplete data. The system recovers each student's ground-truth proficiency tier with 72.5% exact accuracy and retrieves the correct worksheet within the top three results 100% of the time, substantially outperforming a one-size-fits-all baseline (48.5% accuracy) at sub-millisecond retrieval latency. We discuss what this simulation can and cannot establish, and outline a field-trial design as the next step toward classroom deployment.