Arapai: An Offline-First AI Chatbot Architecture for Low-Connectivity Educational Environments
The new architecture enables AI tutoring without internet on legacy, CPU-only devices.
A team of researchers led by Joseph Walusimbi has published a paper on arXiv introducing Arapai, a novel architecture designed to bring AI-powered tutoring to educational environments with limited or no internet connectivity. The system directly addresses the digital inequality exacerbated by most educational AI tools, which require continuous cloud access and modern hardware. Arapai proposes a complementary, offline-first deployment paradigm, enabling curriculum-aligned explanations and problem-solving support in regions where reliable infrastructure is a barrier.
The architecture's core innovation is its integration of locally hosted, quantised language models with an automatic hardware-aware model selection system. This allows it to perform all inference on-device, optimized to run on legacy, CPU-only hardware by keeping models resident in memory. A pilot study in secondary and tertiary institutions under limited-connectivity conditions evaluated the system across technical performance, usability, answer quality, and educational impact. Results indicated stable operation, acceptable response times for instructional queries, and positive perceptions regarding its support for self-directed learning. The work contributes a practical framework for decentralised, infrastructure-resilient educational technology that advances digital inclusion.
- Architecture runs entirely offline using quantised LLMs on CPU-only, low-specification devices.
- Features automatic hardware-aware model selection and pedagogically tiered response control for curriculum alignment.
- Pilot deployment showed stable operation on legacy hardware and positive learner/teacher feedback on educational impact.
Why It Matters
It provides a practical blueprint for deploying resilient, inclusive AI educational tools in bandwidth-constrained regions worldwide.