LectuLibre builds two-pass LLM pipeline to improve AI book translation accuracy

LectuLibre, a platform that translates entire books using large language models like Claude and DeepSeek, developed a two-stage translation system called Direct Translation & Meaning Confirmation to address nuance loss in single-pass AI translations. The team found that a straightforward chunk-and-translate approach struggled with polysemous words, culturally weighted proper names, and idioms that cannot be rendered literally. Their solution first generates a conservative, literal draft translation, then runs a meaning-confirmation step that identifies ambiguous phrases and asks the model to verify or improve them in context. To manage costs on large books, a lightweight NLP-based phrase selector flags only high-risk candidates — such as named entities and words with high translation entropy — rather than re-examining every phrase. The pipeline is built on FastAPI, SQLAlchemy, and PostgreSQL, with translation jobs handled asynchronously via background workers.