The persistent digital divide continues to hinder equitable access to high-quality education, particularly in low-resource regions where internet connectivity, educator capacity, and personalized learning infrastructure are limited. This paper presents an AI-driven assessment system that advances programming education through adaptive feedback tailored to individual learners, even in offline environments. By integrating GPT for generating contextually relevant programming questions, BERT for automatic difficulty classification, and a reinforcement learning engine for dynamic adaptation, the system offers real-time personalization without requiring continuous internet access. In a controlled evaluation, the system generated a curated set of 500 programming questions, with 92% passing automated filtering, and expert raters awarding an average score of 4.4/5 across technical accuracy, pedagogical relevance, and linguistic clarity. The adaptive engine demonstrated strong responsiveness, adjusting question difficulty within 5 to 10 iterations for simulated learners at beginner, intermediate, and advanced levels. Additionally, the system achieved a low average processing latency of 550 ms, ensuring seamless interaction and learner engagement in real time. What sets this work apart is its commitment to educational equity. Designed specifically for resource-constrained settings, the system’s offline functionality allows schools and learning centers with limited infrastructure to benefit from the same level of AI-enhanced learning found in high-resource contexts. By reducing educator workload, scaling individualized instruction, and increasing access to quality programming education, this work exemplifies how AI can be harnessed for social good. It contributes a replicable, scalable model for bridging educational gaps and shaping the future of learning in underserved communities worldwide.