HYBRID EDGE-CLOUD AI FOR RURAL STEM EDUCATION: AN OFFLINE-CAPABLE MOBILE SYSTEM INTEGRATING CLAUDE, GEMINI, AND OLLAMA LLMS TO OPTIMIZE ASSESSMENTS IN LOW-CONNECTIVITY CONTEXTS

Rural STEM education in developing regions faces critical challenges including intermittent connectivity, scarce technological resources, and inadequate access to quality instructional support. This paper presents a novel hybrid edge-cloud AI system designed to optimize assessments and personalized learning in low-connectivity contexts through the integration of multiple large language models(LLMs). The proposed mobile-first architecture combines offline-capable Progressive Web Applications(PWAs) with a distributed AI framework: Ollama’s lightweight LLMs (1.5–3.8B parameters) for edgeprocessing, Anthropic Claude 3.5 Sonnet for deep pedagogical analysis, and Gemini 1.5 Flash for rapidand cost-efficient cloud-based responses. A six-layer technical architecture enables QTI-standard assessments with LaTeX rendering, adaptive testing, and multi-level feedback generation, while maintaining functionality during internet outages through service workers and local storage. Preliminary evaluation at a Colombian rural school demonstrated a 9.6% increase in mathematics scores (Cohen’s d = 0.484), complete elimination of low-performing students (0–35 point range), and 85% student satisfaction with exam preparation support. The hybrid AI system achieved 60–95% feedback accuracy across models while reducing latency from 8–10 days (manual grading) to 15–20 minutes. A comprehensive cost-benefit analysis revealed 99.9% savings over cloud-dependent platforms (e.g., ALEKS McGraw Hill: $1.56 vs. $1,253 per student over five years), primarily through elimination of recurring internet costs. These results validate the technical feasibility and pedagogical efficacy of edge-cloud AI synergies in resource-constrained environments, offering a scalable model for national education frameworks prioritizing digital inclusion and equitable STEM access.