AI agents that write Hindi-medium study material
THE QUESTION
Can AI agents produce syllabus-faithful CBSE study material, much of it in Hindi, within a token budget and behind automatic quality gates?
Founder, Bodham LabsUpdated
What did we build?
- Five working repositories of CBSE study material for classes 3 to 12, in Hindi medium and English, written by AI coding agents (Claude Code, Codex and Antigravity) under shared rules.
- Chapters stored as data: a chapter is a JSON file of typed blocks, and one renderer builds every page.
- Token budgets in the agents' rules: read only what the task needs, keep reads from outside the chapter under about 2,500 tokens per session, and check each section in a small file instead of the whole chapter.
- A stable agent briefing of about 3,500 tokens, kept free of day-to-day status so that prompt caching keeps working.
- Quality gates: a "zero curriculum drift" rule (the textbook is the ceiling), a code linter, and a seven-tier check.
What does the evidence show?
| WHAT | VALUE | SOURCE |
|---|---|---|
| Hindi (Devanagari) tokens vs English, same 12 sentences | 4.55–4.91× on Qwen, Mistral, cl100k · 1.61× on o200k · 1.13–1.30× on Indic tokenizers | our tokenizer benchmark, Oct 2026 |
| Dev-roman vs pure Devanagari, Qwen tokenizer | 623 vs 929 tokens (about a third fewer) | our tokenizer benchmark, Oct 2026 |
| Largest single chapter file | about 98,000 tokens | agent rules, Sep 2026 |
| Chapters failing the newest standard | 28 of 36, being rebuilt | agent rules, Sep 2026 |
| Chapter folders planned, CBSE 2026, classes 8–12 | 206 | project brief, Aug 2026 |
| Commits to the main curriculum repository | 146 | git history, Aug–Sep 2026 |
Where has it led us?
Hindi costs more, and how much depends on the tokenizer: our benchmark measured 1.13× to 4.91× English for the same sentences. An agent that reads a whole chapter can run out of room halfway through a task, so budgets and targeted reading came first, not bigger models.
Writing technical terms in English (dev-roman) used about a third fewer tokens than pure Devanagari on the Qwen tokenizer, so it is both how students speak and cheaper to generate.
A stable briefing is a cost decision. Anything that changes every session lives in a separate status file, so the cached part stays cached.
The quality gate is honest about where we are: most early chapters failed the newer standard, so they are being rebuilt rather than hidden.
What is still open?
- Which quality gates catch the most errors per token spent?
- How do students learn from this material compared with textbook-only study?
NEXT STEP · NOT DONE YET
What would we do next?
Re-run the tokenizer benchmark on a public parallel set (FLORES+ or IN22) and on the real chapter files, then set per-tokenizer budgets in the agents' rules.
Have a problem that needs this kind of research?
Tell us about it