Home
This paper presents an architectural proposal for a modular and adaptive hybrid recommender system designed to improve equitable access to PDF-based educational content for students with visual impairments. The system overcomes the limitations of traditional recommenders, which optimize rankings solely by thematic relevance while ignoring format-related barriers, by implementing a closed-loop approach that prioritizes technical accessibility prior to ranking the educational material. The architecture is structured into four independent layers: automated ingestion via multimodal artificial intelligence, dynamic student profiling, a sequential hybrid recommendation engine, and a semantic presentation interface. The technical viability of the proposal was successfully validated through a functional Python prototype integrated with the Gemini 2.5 Flash model from Google AI Studio, evaluating a repository of real-world documents split between the fields of algebra and programming. Experimental results demonstrated that the system accurately resolves the cold-start problem through an initial technical questionnaire and effectively mitigates the occurrence of false positives in practice, thanks to a reverse feedback loop that automatically updates functional security profiles and restricts resources reported with access flaws.