The AI boom was built in English. The models read it best, the tools assume it, and the documents they were trained on are mostly written in it.
India has 22 languages in the Eighth Schedule of its Constitution and around 1.4 billion people. Only about one in ten told the 2011 census that they speak English at all. For everyone else, a scanned form in Telugu or a land record in Hindi is still something a person has to retype by hand.
We think that is the wrong way round. Language AI should start with the people it has left out.
THATCHED builds that AI, one useful product at a time. We start with reading: turning paper and PDFs in Indian scripts into text people can edit, search and reuse. Then understanding, so documents become data. Then listening and speaking, because many people would rather talk than type. Then an API, so anyone can build on it.
We ship real things first and talk about them second. Every part of this site says whether it is live, next, in research or planned. Our first product is live today.
We price for India, in rupees and by the page. We keep files private and delete them on a timer. We build for phones and slow connections. And we say plainly what does not work yet.
Starting with India's.
Sources: Constitution of India, Eighth Schedule; UN World Population Prospects 2024; Census of India 2011 (10.6% reported English as a first, second or third language).