(This interview has been lightly edited for brevity)

India has redefined what public digital infrastructure can look like. What would a “public infrastructure for language AI” be, and how could India set a global standard in this space?

A public digital infrastructure for language would have multiple layers. The lowest and most important layer is the data layer: datasets for each language. Some languages already have datasets, and for others they still need to be built. In Bhashini’s (India’s national AI platform for multilingual language technology) case, we went into the field and created datasets through brute force: identifying people to speak and then transcribing. That becomes the foundation. Above that is the language model layer, which uses the dataset. The larger the corpus, the better the model you can build. Then comes the platform layer, which looks at how AI services will be consumed, essentially providing the model services as APIs. Above that is the application layer, built by startups, industry or government. Finally, there is the ecosystem or adaptation layer, where each stakeholder interacts with the DPI in their own way because each ecosystem, like each language and dialect, has its own flavour. For national-scale implementation, you plug in the language layer and data layer for each language. For 22 official languages, you would have 22 such plug-ins. If more languages are added, more plug-ins come in, meaning an AI model plug-in and a dataset plug-in. Other countries are beginning to inquire about how we are doing this. Many languages, especially in the developing world, don’t have scripts or enough data to create models. The Digital India Bhashini Division (Bhashini) seems to be a torch-bearer because of this approach.

What is the strongest public-good argument for language AI in India, and how could it deepen inclusion in areas like education, healthcare and governance?

There are multiple divides in society, and many government initiatives have tried to bridge them, for example, the network divide. Language AI brings in digital inclusion because most digital systems were otherwise in English. It also includes people on the other side of the literacy divide, because voice becomes the medium. So language AI helps you include people across digital, language and literacy divides. That allows you to serve society at a much larger scale. This is why it has to be public infrastructure: by definition it must be used by everyone. Once you have that, it naturally crosses sectors: education, healthcare, governance, insurance, banking, regulatory compliance. Many non-compliances happen simply because people don’t know the rules. Language AI can create a better-informed society and make the ease of living much higher. Most conversations about AI in India focus on scale. But language is intimate, local and sometimes hyper-fragmented.

How do you reconcile India’s scale-driven tech approach with the reality that language technologies work best when they reflect how people actually speak?

When we talk about AI at scale, the scale of languages is actually even larger. The classical approach—building a perfect system and then taking it to market—doesn’t work here. With India’s diversity of dialects and language use, you need to build the system while deploying it. If you wait to perfect a product and only then try to serve 1.4 billion people, you will never reach there. Our approach has been different: we build something, deploy it, enable usage, get feedback and improve. There is a pull from people: those whose languages aren’t yet on the platform want them included, and those who are on the platform see the utility and want to help improve it. So we didn’t try to be perfect at the first instance. We tried to be deployable. We demonstrated utility through narrow use cases, got stakeholders interested, and then entered a continuous loop of deploying the system, gathering feedback and making improvements. This approach saves enormous cost in data collection and development, and it builds adoption organically.