French AI Translation Pipeline

Steinsaltz publishes elucidated editions of classical religious texts. Producing a French edition meant translating from scratch, which is slow, expensive, and the reason most editions never get made. The pipeline changes the job from translating to proofreading.

What it does

• Editors approve rather than author. Output arrives as a draft translation for a human to correct and sign off, which is a different and far faster task than starting from a blank page.
• 99% accuracy on the draft, which is the threshold that makes review cheaper than translation. Below it, correcting the machine costs more than doing the work.
• Terminology has to hold across a book. Religious texts carry vocabulary that must be rendered the same way on page 4 and page 400, so consistency across a long document matters more than fluency in any single sentence.

What I built

• The pipeline itself, in Python on GPT-4, running over full texts rather than isolated passages.
• The review interface, in Next.js on Vercel, where editors read, correct and approve.