R7 · HINDI & HINGLISH · TRANSLATIONPROTOTYPEMay 2026

Faithful Hinglish translation with LLMs

THE QUESTION

Can an LLM translate a long English text into readable Hinglish, sentence by sentence, without adding its own commentary?


Founder, Bodham LabsUpdated

What did we build?

  • An experimental agent pipeline that produced a bilingual English + Hinglish EPUB of a 136-page English text. Hinglish here means Devanagari Hindi with English technical terms.
  • Assembly in plain Python standard-library code: pages, glossaries and diagrams.
  • Three versions, built between 1 and 4 May 2026.
  • A style guide: "DD News/NCERT words → English. Coffee shop words → Hindi. Technical terms → always English."

What does the evidence show?

Values exactly as recorded. The source column says where each one was recorded.
WHATVALUESOURCE
Source length136 pagesthe English source text
Versions built3, between 1 and 4 May 2026build outputs

Where has it led us?

  1. Version 1 drifted into commentary. Instead of translating, it described the page: "इस page में author बोलते हैं".

  2. The fix was a strict rule: "The Hinglish layer must be a faithful sentence-by-sentence translation that preserves the original text's meaning." Sentence order stays 1:1 with the original.

  3. Large batches were split so that each one stayed inside the model's output limit.

What is still open?

  • How do we check faithfulness at scale without a person reading every sentence?
  • Do the style rules hold for other kinds of text, such as school chapters or technical manuals?

NEXT STEP · NOT DONE YET

What would we do next?

Write an automatic check that the Hinglish layer keeps the same sentence count and order as the English source.

Have a problem that needs this kind of research?

Tell us about it