Shelton03/EquiRead--Evidence-Repository

Public architecture, demo assets, and technical evidence for EquiRead, an AI-assisted accessibility platform that transforms inaccessible university PDFs into screen-reader-first and voice-readable educational content for visually impaired students in African higher education. # Evidence — Accessible Academic Document Transformation Evidence describes an AI-assisted document-processing system for converting academic material into structured, screen-reader-first accessible content and an accessible PDF delivery format. ## Problem Academic PDFs often preserve visual appearance but lose the relationships a non-visual reader needs: headings may only be bold text, a table may be emitted as unrelated cells, diagrams may have no text alternative, and a two-column page may be read in the wrong sequence. These failures are especially common in scanned modules and handout-style learning material. The system treats accessibility as a document-understanding task. It first reconstructs page structure, then transforms that structure into a linear learning representation, checks the result, renders it as an accessible PDF, and makes the same transformed content available to Vibe Voice playback. ## Architecture The Android client communicates with a FastAPI service that manages document and job state. Processing occurs outside the request path: a document is parsed into a page-level intermediate representation (IR), transformed in overview and section passes, validated and formatted, then rendered for download and voice use. ```mermaid flowchart LR A[Android client] --> B[FastAPI API] B --> C[Document and job state] B --> D[Parser] D --> E[Structured IR] E --> F[Two-pass accessibility transformation] F --> G[Validation and formatting] G --> H[Accessible PDF] G --> I[Vibe Voice input] ``` ## Processing model 1. **Ingest and identify** — validate the submitted format, create a document record, and detect duplicate content before processing. 2. **Recover page content** — extract digital text when available and use OCR only for pages where extraction is insufficient. 3. **Reconstruct structure** — detect blocks, tables, figures, headers, footers, and a reading order for each page. 4. **Build the IR** — retain page number, ordered content blocks, tab …