MathAssist: An Interactive Mobile Application for Mathematics Learners Specially for Developing Countries Using AI, LLM And RAG Based Approach

Mathematics forms the foundation of science, engineering, and technology education . However, many students struggle with mathematical concepts due to limited personalised guidance, insufficient learning resources, and difficulties in understanding problem-solving procedures . These challenges are more severe in developing countries such as Bangladesh,where approximately 16.4 per cent of students face difficulties in mathematics and 53.1 per cent fail to achieve satisfactory
performance.
Recent advances in Artificial Intelligence (AI) have enabled intelligent tutoring systems capable of providing adaptive learning support and automated problem-solving assistance. Nevertheless, most existing solutions, including Photomath and Khan Academy, rely heavily on cloud computing and continuous internet connectivity. Such requirements
limit their applicability in rural and resource-constrained regions of Bangladesh, where internet access and digital infrastructure remain inadequate.
To address these challenges, this paper proposes MathAssist, an offline AI-assisted mobile learning framework that combines Retrieval-Augmented Generation (RAG) and lightweight Large Language Models (LLMs) to provide context-aware and step-by-step mathematical assistance using verified educational content retrieved from a local knowledge base. The proposed memory-efficient Serial RAG architecture enables deployment on low-resource mobile devices while improving the reliability of mathematical reasoning.
The major contributions of this work are summarised as
follows:
• Development of an offline AI-assisted mathematics learning framework with multimodal interaction capabilities
and lightweight execution suitable for devices with less than 2 GB RAM.
• Design of a novel Serial RAG strategy that integrates semantic retrieval and local knowledge indexing with LLM-based reasoning, where experimental benchmarking identifies WizardMath as the most effective mathematical reasoning model among the evaluated LLMs.