Lifting the floor for all: How Google’s Gemini Models are powering Wadhwani AI’s HealthVaani App for ASHA, Anganwadi workers

Access to primary healthcare continues to be an arduous task for a large section of India’s rural population. With the gap between healthcare services and the beneficiaries still wide, 63% of this population looks up to the ASHA (Accredited Social Health Activists) and Anganwadi workers.

More than 2.2 million such community-based frontline health and development workers serve India’s underserved communities. They are usually entrusted with responsibilities of maternal and child care, public health campaigns, disease surveillance, immunisation along with basic health care.

In the latest episode (1) of the Blueprint to Bharat podcast, senior journalist Sonia Singh had a detailed dialogue with Shekar Sivasubramanian, head of Wadhwani AI, and Preeti Lobana, Vice President and Country Manager at Google India on AI’s expanding footprint in India’s rural healthcare and how both Wadhwani AI and Google are partnering to empower the frontline workers.


Singh considers these women to be frontline heroes in communities with limited medical infrastructure. “For decades, families have turned to them with the high-stakes problems that keep parents up at night.”

While the experience, efficiency, and commitment of these workers is nothing short of exceptional, Singh acknowledged their limitations, too. Problems associated with human memory act as a barrier when they face disparate questions around numerous health problems.

As technology races ahead, these workers no longer have to solely rely on their memory when responding to such questions.

AI footprints in rural healthcare

Today, companies such as Google India and Wadhwani AI are using Artificial Intelligence (AI) to find solutions for some of the most complex problems.

Such initiatives that leverage AI in critical areas, especially in India where a consequential chunk of the population is underserved, are significant.

For instance, Google’s Gemini Models have powered Wadhwani AI’s HealthVaani App for the frontline workers. HealthVaani is embedded within the Poshan Tracker, an application already installed on the mobile phones of frontline workers. Through this integration, frontline workers can get reliable, evidence-based answers to the questions they encounter in their day-to-day work, in their native language—for example, guidance on identifying signs of malnutrition, managing common childhood illnesses, supporting maternal and newborn health, or advising families on appropriate nutrition and care.

Sivasubramanian said this initiative is about efficiency, trust, and reliability but stressed on the need to equip these workers with technology to bridge their limitations. “The most important aspect about these workers is that they are integrated into society. They know families, probably have a tenured understanding of health problems in their region and are storehouses of knowledge and competencies, and actually kindness and niceness.”

Powered by Google, Wadhwani AI has partnered with the Ministry of Health and Family Welfare, and the Ministry of Women and Child Development to develop HealthVaani. The app has been successfully piloted with ASHA workers across Uttar Pradesh, Haryana, Meghalaya, Madhya Pradesh, Rajasthan, and Arunachal Pradesh, and with Anganwadi workers in Odisha and Maharashtra.

“It's a huge systemic shift not visible today, but you will see it unfold. It will strengthen the quality, confidence and assurance of healthcare in India,” Sivasubramanian said, “What technology can do is it remembers far better than what human beings can.”

He, however, added that organisations like his have an added responsibility of bringing in credibility to the information provided as it percolates to these workers. “The worker should be able to say that she is providing information through a formal, systematic, and thorough manner. We need to link up accuracy, specificity with the correctness of information to make that possible.”

Explaining the technology, Sivasubramanian said it has a precise, authorised, and confined set of documents or information containing text, images, tables, provided by the medical community. “The design of the app ensures the answer produced after AI rummaging through 2,000 or 2,500 pages comes from the contextually relevant documents.”

He further elaborated that building human feedback loops, through testing and other techniques makes the responses precise, correct, and medically very relevant. “It is the design supported by humans that ensures the accuracy, so that ASHA and Anganwadi workers get the support they need.”

Google saw an alignment with Wadhwani AI’s approach to working with communities at the grassroots level and driving transformation.

“That approach closely aligns with our thinking. Any progress in technology must benefit society,” Lobana said. “It's not about raising the ceiling for a few but lifting the floor for all. When you look at the power of AI, developed and deployed responsibly, it can transform lives,”

Catering to India’s diversity

Google’s AI integration with Wadhwani AI in India is pivoted around linguistic and cultural diversity. For HealthVaani, the intent is to reach people in their native languages so it can understand their local dialects and nuances.

Speaking about the cultural, linguistic, and contextual diversity of India, Sivasubramanian felt each point needed a solution instead of a “one-size-fits-all” solution. “We know how to get there. It takes some time, maybe a couple of years.”

He added the app, currently available in four languages – Marathi, Odiya, Hindi and English – will soon be available in 12 languages.

“You cannot even think of it otherwise,” said Sivasubramanian while thanking Google for backing HealthVaani as it tries to realise some of the dreams to make India a healthier community.

Speaking about Google’s partnership with Wadhwani AI, Lobana said the two companies have had multiple partnerships in the past. “With HealthVaani, the solution needed to be multimodal. Not just text, but voice images, and multiple languages while respecting India's linguistic and cultural diversity. We felt it to be the right deployment for Google AI.”

She said Google’s AI models are designed to be multimodal – combining text, voice and images – along with their multilingual capabilities, making HealthVaani an ideal platform for deployment. “This initiative by Wadhwani AI is designed to empower frontline workers, and underserved communities while respecting India’s linguistic and cultural diversity,”

Elaborating further, Lobana said Google has combined two of their models, which aids ASHA and Anganwadi workers in getting validated health information at their fingertips.

The two models, she said, were the Gemini Flash and Gemini Embedding model.

Gemini Flash is built for super-high effectiveness as it combines speed, cost efficiency, and intelligence and is at the back end when the answers are being surfaced. “The retrieval of information, its moderation and translation are happening, which is very critical because our frontline workers work in environments that are low resource, where the internet connectivity may not be the best. The phones they have may not be on the highest end. So, we have to ensure the model and its capability can function with such low resources to give the best response in the language of choice in a unique cultural context,” she explained.

Meanwhile, Gemini Embedding model provides semantic retrieval by understanding the intent of what is being asked and not just matching exact keywords. For the ASHA and Anganwadi worker, it means she does not have to frame her question in a medical or technical language. She added: “She can ask anything in a conversational tone or in a language of her choice and get the response because it can understand the intent behind the question.”

Ensuring medical accuracy:

Though the technology is multilingual, there are certain linguistic nuances that are hard to decode, leave alone getting medical accuracy in such a scenario. It is important for the technology to understand these nuances and pull out the accurate information based on such scenarios.

Singh opined that getting clinical accuracy right in a local dialect is very hard because in addition to translating medical information, one has to see it in a context that's culturally correct, contextually appropriate, and medically sound.

When asked how HealthVaani gives frontline workers information that is accurate, not just linguistically, but medically, Sivasubramanian said that these technologies work in a manner that goes to a precise, authorized and confined set of documents or information.

“The document can contain text, images, tables, and a variety of expressions the medical community provides as the basis. The design ensures the answer comes from within the contextually-relevant document. “He said, “The measurement is to make sure you've got the right answer,”

Further, he added, as you build in human feedback loops, through testing and other techniques, those answers become precise and medically relevant. So the design and the human support ensure that the technology is responsibly supporting the ASHA and Angwanwadi worker.

The validation of facts while the model is dealing with different dialects, some of which have no standardised written form, no corpus or academic literature is what Google has done efficiently.

“How does Google achieve that?” Singh asked. Lobana explained that Google has an “ecosystem approach” towards solving such problems, with a philosophy of working with the local players, including non-profits, academia, and/or state governments.

“In one such partnership with the Indian Institute of Science, called Project Vaani, we have open sourced speech data for 109 languages and dialects across multiple states and Union Territories to date. We make them available through the government's Bhasini project.”

Interestingly, many of these local dialects and languages do not have any corpus, and are called low-resource languages. Their speakers do not have any access to the digital ecosystem.

“From an inclusion point of view, it is critical for us to design for that. Startups such as Shillong-based MWire Labs used Project Vaani's natural conversational speech dataset to train a voice-recognition system for Garo, a low-resource language traditionally excluded from major AI models.,” she said.

The potential impact

So, how big is the impact?

Such AI integration can significantly alter healthcare on a personal and population level. While the technology equips a frontline worker with immediate validated guidance, and personalised explanations, the millions of conversations taking place can be pivotal in determining disease patterns, emerging health issues, outbreaks, and other aspects of population-level healthcare in India.

This also presents an opportunity to replicate such models for other sectors, including agriculture and education.

Lobana gave the example of AISteth, a startup using MedGemma (Google's collection of open models for medical text and image comprehension), to empower frontline workers by enabling them to carry out lung screenings, heart screenings, and more. “It is a smart stethoscope,” she explained.

In education, too, she said, Google Gemini has a feature called guided learning, that gently guides and cross-questions rather than giving a straight response.

We are still scratching the surface when it comes to harnessing the potential of AI at scale. But, as more explorations in AI unlock newer frontiers, tech companies like Google, along with Indian startups such as Wadhwani AI, through such deep collaborations and partnerships, can solve problems that have plagued our populations for decades.