Educational Middleware Architecture for Explainable Feedback in Braille Learning: Interaction Logs and Large Language Models (LLMs)

This manuscript presents a preprint of an educational middleware architecture for Braille learning based on interaction log capture and explainable feedback generation using large language models (LLMs). The proposed approach addresses a gap in assistive and educational technologies by focusing on the analysis of the learning process rather than only correctness evaluation. The system treats a bidirectional Braille device as a sensor of the learning process, enabling non-intrusive capture of behavioral data such as time per cell, hesitation, and interaction patterns. A middleware layer structures interaction events, computes pedagogical indicators, and provides them as input to a large language model responsible for generating interpretive feedback for teachers. The architecture explicitly separates deterministic analysis from LLM-based interpretation, ensuring transparency, traceability, and alignment with principles of Explainable Artificial Intelligence (XAI). The current results correspond to preliminary technical evaluation and bench testing in controlled conditions, without the collection of data from human participants. A user-based evaluation involving visually impaired students and teachers has been submitted to a Research Ethics Committee (CEP) and will be conducted after approval. This preprint contributes to the integration of assistive technology, learning analytics, and explainable AI in Braille education.