The authors present AI Textbook Auditor, a modular multi-agent pipeline designed for the automated quality assurance of educational textbooks. The system processes textbook PDFs to generate structured reports through two parallel tracks: a Factual and Technical Track that detects inaccuracies using specialized LLM agents, and a Grammar Track that preserves diacritical encoding.
- A Judge Agent filters false positives using domain-specific rules before presenting findings to human reviewers.
- The pipeline supports vision-native page rendering and PyMuPDF text extraction for ingestion.
- In testing on Romanian upper-secondary textbooks, the system identified 56 technical findings in a CS textbook with 62.5% expert-validated precision.
- It also detected 72 findings in a history and social sciences textbook, covering factual errors, ideological bias, and grammar.
The system serves as a triage tool to reduce manual effort in locating candidate issues, though human expert validation remains required before any editorial action.