In July, Vak was accepted into the Sarvam AI Startup Program, with ₹25,000 in credits and access to Sarvam's speech-to-text, text-to-speech, LLM, translation and document-digitisation APIs.
The gap
More than 600 million people speak Telugu or Hindi, and AI systems are starting to make decisions about their loans, health records and job applications. Interpretability research exists. Tooling that shows what those systems are doing does not exist in their language, at their hardware budget, or anywhere near their context. Most of it assumes you speak English, own a GPU and already know what a transformer is.
What Vak is
Vak is an open-source AI literacy and interpretability platform built natively in Telugu and Hindi. The first version is built so that a developer in Hyderabad can run behavioural audits of LLM outputs in Telugu. It is not a translation of an English tool: the audits are written around how Telugu speakers actually frame questions, and around what Telugu language models get wrong.
The Sarvam access matters because the hard part of building for Indian regional languages is the infrastructure layer. Voice input and output in Telugu would have cost us the first year. Now that year can go into the audits.
If you work on Indian regional languages as a developer, researcher or educator and want to collaborate, write to me at mohd.mudassiruddin7@gmail.com.