A Hybrid Text-Extraction and Transformer-Based Classification Pipeline with Asynchronous Fast API Serving for Academic Document Management

Colleges and universities produce assignments, circulars, research papers and notes by the thousand each term, and most of it is still filed and found by hand. Ordinary cloud storage keeps the files but cannot read them. We describe a document management service, built on FastAPI, that extracts text from digital PDFs (PyMuPDF), mixed office formats (Apache Tika) and scanned or handwritten pages (Tesseract OCR after image clean-up), and then assigns each document to one of four academic categories (Assignment, Research Paper, Notes or Circular) with a fine-tuned DistilBERT classifier, with RoBERTa kept as a higher-accuracy alternative. Results are stored in a relational database, searched by keyword, category, uploader or date, and summarised on an administrator dashboard, while JWT authentication and role-based access control protect every endpoint. On an evaluation set of 500 academic files the DistilBERT classifier reached an average accuracy of 93.8%, and processing stayed under 1.5 s per document with five concurrent user sessions. All six functional and security test cases passed. We report the design, the reasoning behind each choice, the measurements, the limits of our evaluation, and how the results sit against three recent studies of OCR deployment.