AI Proofreading
A publishing-grade AI proofreading engine, live on web, Windows, and macOS.
The problem
A book can run to hundreds of thousands of characters. Human proofreading is slow and error-prone. Publishers don’t just need a tool that catches mistakes — they need an auditable, traceable pipeline that fits into the editorial workflow.
What we built
We built a publishing-grade AI proofreading engine with web, Windows, and macOS clients, designed to work alongside mainstream agent workflows. The engine combines multiple models, a rule layer, and an evaluation set: different models handle different error types, the rule layer covers formatting and terminology, and the evaluation set makes every upgrade measurable.
Results
- In a quality comparison across five real manuscripts, professional manuscripts reached a 12.2‰ error rate — first-tier performance.
- After one engine upgrade, issue recall rose from 47% to 56%, while cost fell from ¥2.76 to ¥1.64 per 10,000 characters (−41%).
- The product is live on web, Windows, and macOS, with publishing-grade Track Changes output.
Boundaries
These figures come from a July 2026 comparison on a specified sample and an internal engine experiment. They apply to that sample only; manuscript type, proofreading standard, and model version all affect results.