Sarvam AI Founder Story — Vivek Raghavan & Pratyush Kumar | India's Sovereign LLM Unicorn | UpForge

The India-First AI Vision

In 2023, when the world was fixated on GPT-4 and the race between OpenAI and Google, two researchers in Bengaluru made a quieter but more consequential decision: India needed its own foundation models.

Vivek Raghavan and Pratyush Kumar had spent years at IIT Madras and AI4Bharat — a government-backed initiative to build language technology for the subcontinent. They knew something most Silicon Valley engineers didn't: that 90% of India's population primarily communicates in languages that are massively underrepresented in global AI training data.

Sarvam AI was built on a single premise — that the AI era would only serve India if India had sovereign infrastructure to participate in it on its own terms.

Building India's Own LLM

Sarvam's technical bet was specific and hard. Rather than fine-tuning existing Western models on Indian data, the founders committed to training models from scratch — with architectures that understood the grammatical structures of Dravidian and Indo-Aryan language families.

By early 2024, Sarvam-1 launched: the first LLM trained comprehensively on 22 Indian languages. It was not a translation model. It was a reasoning model — one that could understand context, disambiguate honorifics, and navigate the code-switching between English and Hindi that defines how educated Indians actually communicate.

The early results stunned enterprise customers. Sarvam's voice AI and conversational APIs outperformed global models on Hindi, Tamil, and Telugu by margins that made the product genuinely competitive at enterprise scale.

The Sovereign AI Moment

As the global AI race intensified through 2024, India's government moved decisively. The IndiaAI Mission — a $1.2B national initiative — selected Sarvam AI as its primary model development partner. The implications were enormous: state capacity, government data access, and sovereign deployment rights.

The $41M Series A from Peak XV and Lightspeed followed. Total funding crossed $70M. Valuation crossed $1B.

But what Raghavan describes as the real validation isn't the capital — it's the deployment. Sarvam's models now power government citizen services, agricultural advisory platforms, and healthcare chatbots across India's rural heartland. These are users GPT-4 was never designed for. They speak Bhojpuri, Odia, and Marathi. They use voice, not keyboards.

For them, Sarvam AI isn't a feature — it's the only viable interface.