India may be the world's toughest test case for artificial intelligence and that is exactly why Pratyush Kumar wants to build it here.

Speaking at the NDTV Ind.AI Summit, the CEO and co founder of Sarvam AI said India's scale, languages and diversity make it "a tremendously hard benchmark for what AI needs to get right to work for all of India."

Kumar argued that AI's value loops, where user feedback rapidly sharpens models, now compound in months instead of years. "To innovate, you need to put technology out there and use feedback to make the models better," he said, stressing the need for real world deployment at scale.

Sarvam is betting on Indian languages, optimizing them with a homegrown AI stack and pushing them to serve entire populations. "We are actually world class in this," Kumar said, citing internal benchmarks.

On olmOCR Bench, Sarvam's Vision model scored about 84.3 percent accuracy, ahead of Gemini 3 Pro at roughly 80.2 percent and a ChatGPT vision model at around 69 to 70 percent.

On OmniDocBench v1.5, Sarvam Vision posted 93.28 percent accuracy, again leading competing OCR systems. In word accuracy tests, it recorded 87.36 percent, compared with 82.51 percent for Gemini 3 Pro.

Source:Ndtv