Indian AI-solutions built on Google models being adopted globally: Google Deepmind’s Manish Gupta

Artificial Intelligence (AI) tools built in India using Google’s AI models in areas like medicine, agriculture and language, are now being rolled out globally, Manish Gupta, senior director at Google DeepMind, said on Tuesday. Gupta was speaking about the global tech giant announcing $2 million funding each, for the four AI centres of excellence set up by the government, as well as millions in AI funding for startups and institutions.

"The diabetic retinopathy solution was first developed in India, then also supported screenings in Thailand. Our agricultural models, first launched in India, were launched in multiple countries in Southeast Asia a few weeks ago. Now there are plans to extend that to many other parts of the world," Gupta said.

Google's premier AI research lab, dedicated to creating general-purpose AI that advances science and benefits humanity, Google Deepmind is known for developing Gemma, Google's family of free and open small language models, as well as breakthroughs in AI-led diagnosis of eye diseases, and accurate weather forecasting. Google has now uploaded all 22 Gemma models onto AIKosh, the India AI Mission’s open data and model platform to help developers, Gupta said.


However, Gupta termed Indian industry as 'a huge laggard' in driving research and development. "They spend a very, very small percentage of revenue on R&D. So, the government has started to do its bit, Indian industry needs to do more, if we are to be true leaders. Industry needs to also get its act together," he stressed.

Indic language push

Gupta said Google's Project Vaani has also digitized speech data for 110 Indian languages and made it freely available. A collaborative initiative with the Indian Institute of Science (IISc) and the AI and Robotics Technology Park to collect speech data from Indian languages, it has now collected data of 'zero corpus languages' from over 143 districts.

"For 35 of those languages, this is the first ever speech data. For at least 22 of those languages, this is the first ever digital data known to AI practitioners and researchers," he said.

The tech giant announced a $2 million founding contribution to establish the new Indic Language Technologies Research Hub at IIT Bombay. It will also provide $50,000 each to startups Gnani.AI and CoRover.AI, currently leveraging Gemma to build Voice AI models and e-governance–focused models serving Indic language solutions.

In the health segment, Google on Tuesday announced it will provide $2.5 million to support the non-profit research institute Wadhwani AI in piloting HealthVaani, a large language model (LLM)-based conversational AI assistant, as well as $2 million to help develop and deploy Garuda, a new Indian language model for agriculture.

Referring to the 2024 chemistry Nobel was awarded to Google Deepmind CEO Demis Hassabis and others for developing an AI tool to predict protein structures called AlphaFold, Gupta said the predictive structures of all 200 million proteins from humankind mapped by the tool are now being used by over 3 million researchers across the world for a broad range of application. This includes Indian researchers working on drug discovery or making food like soybean resistant to disease or designing enzymes that would make degradable plastics, he added.

Gupta stressed that AI has shown the potential for dramatic impact across material science, material discovery, weather forecasting to clean energy. "Google has supported nearly 1,000 years of PhD-level research work across more than 25 leading research institutions, including premier institutes," he said.

Research estimates that AI-assisted screening tools could save the Indian public roughly $4.7 billion yearly, which is approximately 12% of India's total out-of-pocket healthcare expenditure, Google said in a statement.

The tech giant also announced funding worth $400,000 to support new collaborations to leverage MedGemma, its family of open, specialized AI models designed for healthcare, to build India’s first health foundation models. It is also working with the National Health Authority (NHA) to deploy its advanced AI to convert millions of fragmented, unstructured medical records such as doctor's clinical and progress notes into the international, machine-readable FHIR standard. Over 400,000 NHA-registered health facilities, including hospitals, clinics, and diagnostic labs, are also being systematically brought on Google Maps and Search.

Utilizing Google's Open Health Stack, a collection of open-source tools for next-generation digital health solutions, non-profit Khushi Baby has successfully conducted over 35 million tuberculosis screenings in Rajasthan.