The naive approach to multilingual education AI is to reason in English and translate the output. It produces text that is grammatically fine and pedagogically flat.
Why translation falls short
Explanations carry structure — the order concepts are introduced, the analogies chosen, the terms assumed familiar. That structure is language- and culture-specific. Translating the surface leaves the English scaffolding underneath.
Code-mixing is the real classroom
Students do not speak in one language. A doubt arrives as Hindi grammar carrying English technical terms, and the answer needs to come back the same way. Handling code-mixed input was a requirement, not a refinement.
Script and terminology
- Terminology consistency. A technical term must match the one in that state’s textbook, not a dictionary alternative.
- Script rendering. Devanagari, Bengali, Tamil and Gujarati each carry typographic requirements that break naive text handling.
- Register. The formality a teacher uses differs from what a student expects from a study companion.
Where teachers close the loop
Every correction a teacher makes to regional-language output is a training signal. The Marathi and Gujarati layers improved measurably over one term, largely on the back of classroom corrections rather than additional model work.