Translating into Hinglish with an LLM: the rules that stopped the drift

What do we mean by Hinglish?

Hindi grammar and everyday Hindi words, written in Devanagari, with English kept for technical terms and for the words people actually say in English. It is how many Indian classrooms, offices and families talk. A sentence like “Battery circuit में current को push करती है” is normal speech in a Delhi physics class. A fully Sanskritised version is correct and hard to follow.

That is why it matters to us. We explain things this way in class, and our Hindi and Hinglish AI work is built for people who speak it.

What did we build?

In May 2026 an experimental agent used an LLM to turn a 136-page English text into a bilingual EPUB: each English passage with its Hinglish translation. We built three versions between 1 and 4 May.

How did version 1 drift?

Version 1 stopped translating and started describing. Instead of rendering a sentence, it would write lines like this:

That is commentary, not translation: “on this page the author says…”. Language models are trained to be helpful, and with a loose instruction, summarising feels helpful. For a reader it is a trap. They can no longer tell the author's words from the model's.

What rules stopped the drift?

Two rules did most of the work. The first is about faithfulness:

In practice that meant strict 1:1 sentence order. One Hinglish sentence for each English sentence, in the same order. No merging, no skipping, no introductions. The second rule is about which words stay English:

The idea: if a Hindi word is one you only hear on the news or read in a textbook, use the English word people actually say. If it is a word you would use over coffee, keep it in Hindi.

What does a faithful sentence look like?

Here is one sentence, written for this note, in three versions:

The faithful version keeps the order, the meaning and the technical words. It adds nothing.

How would you check a long translation at scale?

1. 01Count sentencesEach page should have the same number of sentences in both languages.
2. 02Search for meta phrasesLines like “इस page में” or “author कहते हैं” are almost always drift.
3. 03Keep batches smallSplit long sections so the model never runs out of output mid-page.
4. 04Have a person read samplesPick random pages and compare them with the source, sentence by sentence.

What this does not prove

• No quality score yet. We have not had human reviewers score the translation.
• One text, one setup. Another subject or another model may drift in other ways.
• A prototype. It is not a claim that LLM Hinglish is ready to publish without a human editor.

Questions