RLHF at scale across 11 Indic languages

Multi-language preference data + reward model debugging that lifted preference win-rate by 18 points. — full narrative will be published once the case-study body field is filled in Sanity. The layout below is production-ready.