Krishna Permi recalls how it all began. “I wanted to help my kid with Kannada homework,” says the 35-year-old independent developer from Bengaluru. It was this simple task at home that saw Permi develop Akshar, a free keyboard app that supports 21 Indian languages.
Permi says he had been helping his son practise Kannada at home. To create reading exercises and notes for his child, he needed a way to type comfortably in Kannada on his iPhone. While he had learned to read and write the language by hand growing up in northern Karnataka, typing it digitally did not come naturally.
Permi ended up doing what most people in his position do: he opened Google Input Tools in a browser. The web tool, which hasn’t changed much since its launch, lets him type phonetically in English and see Kannada characters appear on screen. “But whenever I used it on my phone, it was still not optimised for mobile,” he says. It required an internet connection, and it felt, in his words, “still stuck in that era”.
So he built an app himself. He took the Google Input Tools API, wrapped it in a native iPhone app, and used it for about a year. Eventually, he decided to put it on the App Store. That’s when a small but consequential detail surfaced. When Apple asked developers to declare what data their app collects, the privacy nutrition label, Permi realised that even though he personally wasn’t storing anything, every transliteration query was still going out to an external server.
“This didn’t feel right because I had no control over what data was being sent to Google. The keyboard also required an internet connection to work. At the same time, I started looking for a fully private, offline, on-device alternative,” he says.
When it comes to the software stack, keyboards occupy an unusually sensitive position, as unlike most apps, they sit between the user and everything else, logging in every character typed. Third-party keyboard developers have access to an optional ‘full access’ mode that enables additional features but also, in theory, enables data transmission. This is an area where user trust is hard to earn and easy to lose.
“People use keyboards for a lot of personal stuff such as OTPs, credit card numbers, all of that,” he says. “I wanted to give that assurance that I am not collecting anything.”
Permi’s quest for a privacy-preserving alternative led him to AI4Bharat, a research initiative based at IIT Madras that focused on natural language processing for Indian languages. Among their open-source releases was IndicXlit, a transliteration model trained across 21 Indian languages that converts phonetic English input into native scripts.
Permi shared that while the IndicXlit was being used in institutional contexts, he could not find any consumer-level apps that had integrated it. This is when he decided to build one. “I felt a little happy and emotional too because when you keep hearing about models from somewhere outside India being put to use at a consumer level, and then you find something that you can integrate into a keyboard without sending any information to any server or API, doing everything on-device, it feels like a breakthrough. That felt very special,” he shared.
When asked how the AI4Bharat IndicXLit model works entirely on-device, Permi admitted that there were some challenges on the path. The keyboard extensions in iOS have strict memory limits, around 50 to 75 MB. The IndicXlit model, as released, came in at roughly 350 MB and carried approximately 15 million parameters.
To fit it inside a keyboard, Permi had to compress it significantly without degrading its output quality. The technique is called quantisation, which is essentially reducing the numerical precision of a model’s weights so they require less memory while preserving most of their behaviour. Less precise, but still meaningful, and far lighter.
“You can compress it without losing any accuracy,” he said. “That’s when I could put it on the keyboard and make sure the size was actually usable.”
The result was Akshar, a keyboard that runs entirely on-device, requires no internet connection, and collects no user data. It supports 21 Indian languages, including several that don’t appear in competing tools such as Konkani, Bodo, Kashmiri, Dogri, Maithili, Manipuri, Sanskrit, and Sindhi. For comparison, Gboard supports six Indian languages for transliteration; Apple’s own system keyboard supports 11.
For speakers of languages like Kashmiri or Bodo, the absence from mainstream keyboards isn’t incidental; it reflects a broader pattern in technology products, where support tends to follow user base size and business viability. “When any large organisation wants to add a feature, they think, ‘Is there enough of a user base?’ ‘Will the traction justify the effort?’” says Permi. “As an independent developer, I can just ship it, keep improving it, and put it in people’s hands.”
Ahead of the interview, we tested the keyboard. While it offers an easy segue into scripts in different languages, a few nicks persist. When it comes to accuracy, Permi acknowledges that it is still a work in progress. He shared that the model produces multiple suggestions for each input and that he spent a considerable amount of time building a custom dictionary layer to push the most common words to the first slot.
Permi revealed that he tests every language with batches of 20 to 30 real-world snippets, including mixed-language phrases of the kind people actually type. Languages like Kashmiri and Manipuri, which share characters with neighbouring scripts, have required particular attention. For those, he keeps updating the dictionary based on user reports and his own testing.
There are gaps. Numbers don’t yet integrate cleanly with transliterated text; typing something like ‘Drishyam 3’ in Hindi can produce unexpected results. Autocorrect within Akshar is currently far from Apple’s standard, which Permi says is the next feature he’s working on: making the keyboard reliable enough in English that users don’t feel the need to switch back to the system keyboard.
Akshar launched on the App Store a few weeks before this interview. When asked about the reception, Permi revealed that the feedback came from quarters he hadn’t anticipated. A user in Maharashtra wrote to say he was using it to send WhatsApp messages to his business customers in Marathi. Further, when Permi posted about the app on Reddit, a thread of responses came in from second-generation Indians in the United States—people who grew up speaking Indian languages at home, can read and write them, and have spent their whole lives typing on English keyboards. “They said that this is exactly what they were looking for,” he recalled.
The app is free, and Permi says it will remain free. The economics are simpler than they might appear: because everything runs locally via Core ML, there are no server costs, no API fees, and no scaling infrastructure. “I’m not having any hosting costs or server costs,” he says. “Everything is taken care of by Core ML, running everything locally.”
When asked if he has plans to bring the keyboard to Android, Permi said that it is under consideration. He works with Swift, an intuitive open-source programming language developed by Apple for building apps across iOS, MacOS, watchOS, and tvOS. Official support for developing Android apps using Swift was announced in October 2025.
Permi shared that he wants to bring Akshar to Android users eventually—not least because a large majority of Indian smartphone users are on that platform. He is also keeping a close watch on AI4Bharat’s research output for new language models. Gondi, an officially recognised Indian language, is not yet supported by IndicXlit. When a model becomes available, he said, he will integrate it.
When asked what he wants to see change at scale, the keyboard, he said, is a small step toward a broader vision. He sees technology as a way to help people interact, learn, and collaborate without being limited by the language they speak.
