India’s tech bros’ new fad is Indic AI — Krutrim to AppsForBharat. The answer to ChatGPT

The indigenous AI’s idea is rooted in the Hindu ecosystem’s version of history, wherein India was the epicenter of development, up until the onslaught of invaders.

Illustration by Prajna Ghosh | ThePrint

Start-up founder Prashant Sachan visited San Francisco, the axis of the global start-up ecosystem, only to be exposed to the filthy and criminal side of the city. There was shouting, there was crying, and of course, there were guns. It became the affirmation he had been searching for and the perspective he needed –– there’s a lot to be grateful for in India. SriMandir, an app dedicated to the ‘devotional needs’ of a billion Hindus, is his way of giving back to India and Indians.

The India stamp on the brave new artificial world is being inked in Bengaluru. And with the government’s demand to BuildForBharat, tech-bros are rushing to colonise and Indianise content and AI applications. On X, they are pushing for India’s technological acceleration, with e/acc, a symbol for an internet-born “philosophical movement” that puts tech at the forefront, in their bios –– all through the products they create. A number of Indian entrepreneurs are convinced they can revive their nation’s lost glory, and luxuriate in its newfound success. ‘Building for Bharat’ is now a mantra that’s not only creating tech-based solutions, but is also a cultural project. By building Hindu-first apps and ‘indic-LLMs’ (large language models), founders are forging ‘indic’ narratives.

They range from practical and task-oriented assistants like Krutrim, to digital disseminations of Hinduism. From AI-based chatbots moonlighting as gods to religious texts doling out advice to the faithful. It’s a burgeoning field with a dizzying variety. From apps catering to farmers’ needs to those claiming to enable financial literacy in rural India.

IIT engineers, spread between Bengaluru and Chennai, are creating Indic LLMs –– models that are not only rooted in “Indian culture and sensibilities”, but are in languages like Hindi and Tamil.

Indian bot, Indian narrative

About a month before the beta-launch of Krutrim, the Sanskrit word for artificial, founder Bhavish Aggarwal encapsulated his ethos in a single tweet. He posted two side-by-side screenshots, one of OpenAI’s ChatGPT, and the other of Krutrim. Both models were answering the same question: “Was India a country before the British Raj?”

GPT4, ChatGPT’s latest avatar’s answer is a lukewarm ‘no’. “Before the establishment of the British Raj, the region was not a single unified country in the modern sense, but a collection of various kingdoms, empires and principalities,” the generative AI bot goes on, adding that “The concept of India as a unified nation-state emerged more clearly during and after the British colonial rule.”

Krutrim’s answer is a resounding ‘yes’. “Yes, India was a country before the British Raj. It had its own political system, culture, religion, and tradition. However, after the arrival of the British, India’s political structure changed drastically due to the colonial policies implemented.”

“This is why we need our own AI,” Aggarwal declared on X.

The indigenous AI’s idea is rooted in the Hindu ecosystem’s version of history, wherein India was the epicenter of progress and development, up until the onslaught of invaders –– Mughals, British, sometimes the Congress. This history first calcified through WhatsApp forwards, and is now receiving an additional layer of validation through AI.

IIT engineers, spread between Bengaluru and Chennai, are creating Indic LLMs –– models that are not only rooted in “Indian culture and sensibilities”, but are in languages like Hindi and Tamil. The founders of SarvamAI, Vivek Raghavan and Pratyush Kumar, also launched AI4Bharat, a Nandan Nilekani-backed research lab at IIT Madras. A problem that afflicts Indic models is that there isn’t enough data for them to consume. AI4Bharat aims to end this –– by collecting data from over 700 Indian districts.

There’s piles of data lifted from ancient Indian texts and history textbooks to feed indicLLMs. But there are also available less hallowed options and ways of deploying the bot –– the Krutrim website recommends using it for dhokla recipies, sari styling, kurti styling

The Ambanis have also thrown their hat into the ring. Mukesh Ambani is due to launch Hanooman, a collection of Indic LLMs trained on content in 22 languages.

And behind this attempt is the prevailing belief system that Global LLMs cannot be moulded towards the Indian psyche.

“The true history of the world is different from what is available through global sources. They have documented history in a certain manner. According to them, India didn’t exist as a nation state. But what we have in our resources is different. Their resources tend to represent Indian history in a simplistic manner,” says Ravi Jain, Head of Strategy at Krutrim. “Their corpus of data on how America came into being, would be far more extensive.”

Generative AI platforms like ChatGPT, Google’s Gemini, Musk-backed Grok are global LLMs trained on widespread data. Krutrim, while trained on global data, is fed on a diet of Indian history textbooks and websites.

There’s piles of data lifted from ancient Indian texts and history textbooks to feed indicLLMs. But there are also available less hallowed options and ways of deploying the bot –– the Krutrim website recommends using it for dhokla recipies, sari styling, kurti styling, information on Indian cultural festivals, and sandwiched in between is the option for a Bengaluru travel guide.

But start-ups can’t deliver alone. That’s where the government needs to step in, says Jain. It’s not enough to digitise ancient texts and preserve age-old manuscripts for their historical significance –– their data also needs to be made publicly available, and be online, particularly because indicLLMs need to be trained on them.

The tech narratives need to reflect this “pure, idyllic Bharat” too.

It’s an exercise in instilling pride, cultivated because engineers received rude awakenings in the first world. “Go to downtown SF, there are drug addicts, guns, and so many issues. At this point in time, our country is a lot safer, a lot better. And that’s because our cultural fabric is so important,” says Sachan. And that why he feels building for India is so important.

Capitalising on the trend

Prashant Sachan started AppsForBharat with a mission in mind –– to revitalise the temple experience for over a billion Indians. There’s no cash in wallets, no pens on paper, and barely any books on shelves. Temples required a convenient digital overhaul as well, and that’s when he stepped in with the SriMandir app.

Sachan and Chamaria, investor-turned-builder at AppsForBharat, attribute their success to a cultural shift. According to them, with this government in power, Indians are far more expressive about their religious identity. And thech bros have capitalised on the moment.

Sachan, an IIT-Bombay graduate, wants to fill a void for his countrymen. He looked no further than his family –– he wanted to build for the average Indian.

“My family’s a representation of what mass India looks like. There are no resources, no infra. But there’s hope. You work hard, hope for the best. This is why we’ve done well as a nation,” says Sachan.

According to him, it’s all about praying. He saw it firsthand during Covid-19. He had rushed back home, was admitted at a crowded hospital forced to contend with an influx of disease and the absence of health infrastructure. But at least people had faith.

SriMandir fulfills dual requirements. First and foremost, is the digital shrine. It’s user friendly to the tee. You can select the god of your choosing, modify them down to their hair and skin colour –– depending on the avatar you’re used to praying to. As opposed to visiting your temple of choice, you can access god through the app and go about your daily prayers. There are also astrologers that can be consulted, an exhaustive list of bhajans and devotional songs to skim through. There’s a bunch of “Hindu literature” to peruse, and digital darshans to embark on. There’s also an algorithmically induced ‘god of the day’. It’s a fully contained Hindu experience, complete in and of itself.

This confirmation of India’s cultural supremacy is also Krutrim’s governing philosophy. Jain compares global alternatives like ChatGPT and its ilk with consuming news only through foreign newspapers like Washington Post and The New York Times. They cover India –– but there are limitations. And it doesn’t always lead to a positive image.

In fact, according to Jain, it’s a pretty ominous picture. “In no time, AI will pervade every single user interface –– how we interact with companies, with every application. For example, when you seek medical advice [you will be interacting with an AI model,” he says. “In the absence of an Indian model, we will not even know how or when, but we will be completely westernised. We will lose sight of who we were.”

SriMandir fulfills dual requirements. First and foremost, is the digital shrine. It’s user friendly to the tee. You can select the god of your choosing, modify them down to their hair and skin colour –– depending on the avatar you’re used to praying to. As opposed to visiting your temple of choice, you can access god through the app and go about your daily prayers.

AI, the window to the world

What is data today, will be shaping perceptions tomorrow. Data is the new food. That’s why the developers of Indian AI are cautious about what they consume. That’s where IndicLLMs come in. A research area that’s currently expanding and looks at the long-term effects of AI usage perception, says Jain.

Indian problems need Indian solutions. “Tech needs to be put to service,” says Sachan. Krutrim, for one, has been trained on two trillion tokens, sub-words used in ‘indic’ languages. Essentially, a word is split into parts in order to train a model. AI4Bharat has built a number of data-sets, catered towards a variety of fields — students, governments, academics –– with names like IndicCorp, Shrutilipi, Kathbath, Indic4Trans, among others. Each contains thousands of hours of recorded speech, spanning both languages and districts.

“From time immemorial, we have been the source of knowledge and information for the world. Steve jobs came here for enlightenment. Everyone takes inspiration from us. It needs to continue like this. We’re a very interesting country,” says Jain. “We’re complicated.”

Effectively, if India doesn’t contribute to the larger AI space, the world is missing out. “We’ll have nothing to offer. Like the media is a window to the world, so is AI.”

However, there are multiple Indias, a potpourri of cultures and frames of reference. The only way to build “an ideal model or product” is to imbue it with as much data as possible. “It should use all data that is Indian,” says Jain. If the need of the hour is to encompass what India is through AI, the net needs to be spread as wide as possible. “In the true Indian sense, the perfect model should give all perspectives.”

And these perspectives aren’t necessarily written in English. The problem is the absence of “high quality and diverse indic-language content at scale.” That’s where open-source data-sets, such as those created by AI4Bharat, enter the scene.

The market

The ultimate aim is to serve the needs of the ever-growing and aspirational Indian community at large, whether it’s as knowledge sources, or in the case of SriMandir, an open-armed embrace of religion and the corresponding  market.

It’s why SriMandir has done particularly well in serving the needs of the Indian diaspora. There are 52,000 users in the US, second only to India. When Sachan visited the US, he met many of his overseas consumers –– Hindus who were grateful they didn’t need to travel to their local temples, and because they could digitally donate as well. A number of users he met were from Sunnyvale, California, an American city known for its Indian IT community.

“My life has changed. The app keeps me joyful the whole day,” writes a user on the app’s comment section, one in a sea of the community’s favourite internet-easy abbreviation: JSR for Jai Shri Ram. There’s also JB for Jai Bajrangbali. Instead of trudging to the temple each day, or not praying at all, he’s now doing everything digitally.

“When we started at the height of the pandemic, no one imagined that the temple ecosystem could be changed like this,” says Ayush Chamaria, investor-turned-builder at AppsForBharat. He’s an IIM Ahmedabad graduate who is moderately religious.

Those employed at AppsForBharat aren’t necessarily religious, or Hindus. “If I’m working at Meesho, that doesn’t necessarily mean I’m interested in clothes,” says Chamaria. SriMandir has over 10 million downloads of users, and in 2021, raised $10 million in funding from VC giants, including Sequoia Capital India and Matrix Partners India.

Instead, it has spawned a number of temple-based apps such as Sutradhar and MyMandir, though none operate on the scale SriMandir does. There are also now astrology apps, like AstroTalk, which connects users to India’s “best astrologists.”

Additionally, chat bots powered by OpenAI’s now-public API which permit users to speak to their favourite gods, are also in vogue. There are multiple iterations of ‘GitaGPT’ — described as ‘Bhagvad Gita inspired AI’. One bot claims to “leverage the wisdom of the Bhagavad Gita to offer modern insights and guidance.”

“The political tailoring definitely helped us, the app has worked because the BJP is in power. For example, the inauguration of the Ram Temple has definitely helped us. Even so, devotion and religion have existed as needs forever,” says Chamaria. “Now people know, it’s not a bad thing to project their religion.”

Arguably the country’s biggest ever projection of religion took place on 22 January, when the Ram Temple was consecrated at Ayodhya. Lakhs of Hindus wanted to be there, but were not, following the Prime Minister’s humble request –– the temple should not be crowded, but diyas should be lit at home.

Naturally, there was extreme disappointment. But that’s when SriMandir entered the scene. With the app as facilitator, 50,000 diyas were lit on the banks of the Saryu river. Users received digital certificates, as well as pictures to show that the deed had really been done — part of SriMandir’s ‘chadawa’ project.

“I won’t be present in Ayodhya on the 22nd, but I take pride in @SriMandir_App for facilitating digital access to Ayodhya for countless devotees like myself. Let Ayodhya radiate its brilliance with our collective participation, no matter our location!” wrote an ecstatic Shiv Prasad Barik on Twitter, sharing a picture of his certificate.

One peril of technology, particularly with AI, is that as models learn, they also become unpredictable. They improve, but there’s no certainty of accuracy, even if what’s interpreted as accurate is culturally denoted

Plugging the memory gap

Krutrim and other generative AI models have a memory that is still being built. There are gaps. Krutrim says it was the West Indies that won the 1983 World Cup. And was unable to answer who killed Mahatma Gandhi. Other AI models are also known for, as it has come to be known in common parlance, ‘hallucinating’. The fledgling model has made damning mistakes that X users were far from pleased with.

But Jain says it’s constantly learning. “It’s just a matter of time.”

One peril of technology, particularly with AI, is that as models learn, they also become unpredictable. They improve, but there’s no certainty of accuracy, even if what’s interpreted as accurate is culturally denoted, such as in the case of India’s existence prior to the British Raj. And with desi generative AI apps such as Krutrim being trained on OpenAI’s datasets, there is overlap in the responses vis a vis ChatGPT. Answering the question whether India was a country before the British Raj, Krutrim has also responded: “no, India wasn’t a nation-state in the traditional sense,”  just like its American counterpart. It’s a work in progress.

“A lot of AIs are created on open-source data, or fine-tuned data-sets, which helps them solve problems. These are instructions given to the base model. There’s not always quality control. These are blows on the chin we’ll have to take,” says Jain.

Open source data is that which is publicly available, but in order for a model to be able to delve deeper, data needs to be more specific, which is where fine-tuning data sets come in –– the model is further trained on domain specific data.

The idea of building for the nation is also measured by way of usefulness — how beneficial an app can be, the ways in which it can upgrade average life. What makes Indian AI models beneficial is that they also have voice-driven interfaces, making them particularly suited to the Indian citizenry, many of whom aren’t as comfortable with text. SarvamAI announced it was working with Microsoft to make their Hindi indic LLM available on Azure, Microsoft’s cloud computing service.

As of 2024, the country has 751.5 million internet users, which amounts to 52.4 per cent of the population having internet access. ‘Building for Bharat’ implies catering to them, to building almost ubiquitous, socially pervasive tech infrastructure like UPI.

But, in order to build for Bharat, being well-versed in ‘Indian’ culture is a must. “There are so many apps for Muslims and Christians, but none for Hindus. We’re an underserved community,” says Sachan.

(Edited by Anurag Chaubey)