Case StudiesScholia — AI editing & translation for an academic journal

15 hours saved per peak week.

Project
Scholia
Client
Humanities research institute
Industry
Publishing / Humanities
Role
AI Engineer (Design + Build)
Context

A small editorial team preparing ~16 articles a week across languages, with translation, grammar and reference work scattered over separate tools — and 33% of articles bouncing back for rework.

Challenge

The team prepared on average 16 articles a week ahead of conferences and volume releases. Editing happened in Word, but translation, grammar, and reference work lived in separate tools and subscriptions. Generic translation tools did not respect the discipline's terminology, the journal's style, or its formatting conventions, so small issues accumulated and reached review: 33% of articles came back needing rework, and during peak weeks the work spilled into evenings and Saturdays.

Approach

I built Scholia — AI editing and translation software shaped around the way the team prepares each article.

Translation runs as a translator → reviewer → human editor chain. A translator agent translates the article segment by segment, using a custom glossary tuned to the discipline's terminology and the journal's style, and leaves a short note wherever it made a judgment call. A reviewer agent then reads the whole translation for consistency of terms and style and corrects what the translator got wrong. The human editor approves each segment into the working draft, edits it, or leaves feedback. Feedback goes back to the reviewer with the segment's full history, and anything it changes returns to the human editor for approval.

During the main edit it flags grammar, expression, and style issues inline, cross-checks references and claims against the team's uploaded sources, and answers questions about the document through a cited Q&A assistant. The editor stays in the loop on every change.

Results & impact

The software saved 15 hours per peak week, without adding headcount. Bounce-back dropped from 33% to 9%, and peak weeks no longer required evenings and Saturdays. The editors got their focus back, with always-on AI handling the repetitive, error-prone work.
15 hours*
Saved per peak week
9%*
Bounce-back rate, down from 33%
1
Instead of 4 separate tools

* Estimated impact based on workflow analysis and pilot observations.

How it works

  • NLP
  • Agentic editing
  • Multi-agent translation
  • Human-in-the-loop
  • Custom glossary

Tech stack

  • FastAPI
  • LangGraph
  • Next.js
  • PostgreSQL
  • Azure

Have a similar problem?

Let's have a virtual coffee. We will talk about your workflow and see whether AI is a sensible way to improve it.

Book a Free Intro Call →