How a father’s solution to son’s Kannada homework became free keyboard app for 21 Indian languages | Technology News

Permi ended up doing what most individuals in his place do: he opened Google Enter Instruments in a browser. The net instrument, which hasn’t modified a lot since its launch, lets him sort phonetically in English and see Kannada characters seem on display. “However each time I used it on my telephone, it was nonetheless not optimised for cell,” he says. It required an web connection, and it felt, in his phrases, “nonetheless caught in that period”.

So he constructed an app himself. He took the Google Enter Instruments API, wrapped it in a local iPhone app, and used it for a couple of 12 months. Ultimately, he determined to place it on the App Retailer. That’s when a small however consequential element surfaced. When Apple requested builders to declare what knowledge their app collects, the privateness diet label, Permi realised that regardless that he personally wasn’t storing something, each transliteration question was nonetheless going out to an exterior server.

“This didn’t really feel proper as a result of I had no management over what knowledge was being despatched to Google. The keyboard additionally required an web connection to work. On the identical time, I began searching for a totally non-public, offline, on-device different,” he says.

Keyboards and privateness

With regards to the software program stack, keyboards occupy an unusually delicate place, as in contrast to most apps, they sit between the consumer and the whole lot else, logging in each character typed. Third-party keyboard builders have entry to an non-obligatory ‘full entry’ mode that allows extra options but in addition, in idea, allows knowledge transmission. That is an space the place consumer belief is tough to earn and straightforward to lose.

“Folks use keyboards for lots of non-public stuff reminiscent of OTPs, bank card numbers, all of that,” he says. “I wished to present that assurance that I’m not amassing something.”

Permi’s quest for a privacy-preserving different led him to AI4Bharat, a analysis initiative based mostly at IIT Madras that centered on pure language processing for Indian languages. Amongst their open-source releases was IndicXlit, a transliteration mannequin skilled throughout 21 Indian languages that converts phonetic English enter into native scripts.

Permi shared that whereas the IndicXlit was being utilized in institutional contexts, he couldn’t discover any consumer-level apps that had built-in it. That is when he determined to construct one. “I felt slightly blissful and emotional too as a result of if you maintain listening to about fashions from someplace exterior India being put to make use of at a shopper stage, and then you definately discover one thing which you can combine right into a keyboard with out sending any info to any server or API, doing the whole lot on-device, it appears like a breakthrough. That felt very particular,” he shared.

Engineering the answer

When requested how the AI4Bharat IndicXLit mannequin works totally on-device, Permi admitted that there have been some challenges on the trail. The keyboard extensions in iOS have strict reminiscence limits, round 50 to 75 MB. The IndicXlit mannequin, as launched, got here in at roughly 350 MB and carried roughly 15 million parameters.

To suit it inside a keyboard, Permi needed to compress it considerably with out degrading its output high quality. The approach is known as quantisation, which is basically lowering the numerical precision of a mannequin’s weights in order that they require much less reminiscence whereas preserving most of their behaviour. Much less exact, however nonetheless significant, and much lighter.

“You possibly can compress it with out dropping any accuracy,” he mentioned. “That’s after I might put it on the keyboard and ensure the scale was really usable.”

Akshar permits customers to sort phonetically in English and immediately convert textual content into Indian languages, all whereas working totally on-device. (Categorical Picture)

The outcome was Akshar, a keyboard that runs totally on-device, requires no web connection, and collects no consumer knowledge. It helps 21 Indian languages, together with a number of that don’t seem in competing instruments reminiscent of Konkani, Bodo, Kashmiri, Dogri, Maithili, Manipuri, Sanskrit, and Sindhi. For comparability, Gboard helps six Indian languages for transliteration; Apple’s personal system keyboard helps 11.

For audio system of languages like Kashmiri or Bodo, the absence from mainstream keyboards isn’t incidental; it displays a broader sample in expertise merchandise, the place assist tends to observe consumer base measurement and enterprise viability. “When any massive organisation desires so as to add a function, they suppose, ‘Is there sufficient of a consumer base?’ ‘Will the traction justify the trouble?’” says Permi. “As an unbiased developer, I can simply ship it, maintain bettering it, and put it in individuals’s arms.”

Forward of the interview, we examined the keyboard. Whereas it provides a straightforward segue into scripts in numerous languages, a number of nicks persist. With regards to accuracy, Permi acknowledges that it’s nonetheless a piece in progress. He shared that the mannequin produces a number of ideas for every enter and that he spent a substantial period of time constructing a customized dictionary layer to push the most typical phrases to the primary slot.

Gaps and journey to accuracy

Permi revealed that he checks each language with batches of 20 to 30 real-world snippets, together with mixed-language phrases of the sort individuals really sort. Languages like Kashmiri and Manipuri, which share characters with neighbouring scripts, have required specific consideration. For these, he retains updating the dictionary based mostly on consumer studies and his personal testing.

There are gaps. Numbers don’t but combine cleanly with transliterated textual content; typing one thing like ‘Drishyam 3’ in Hindi can produce surprising outcomes. Autocorrect inside Akshar is at the moment removed from Apple’s customary, which Permi says is the following function he’s engaged on: making the keyboard dependable sufficient in English that customers don’t really feel the necessity to change again to the system keyboard.

Constructed utilizing AI4Bharat’s IndicXlit mannequin, Akshar helps 21 Indian languages and works offline with out sending consumer knowledge to exterior servers. (Categorical Picture)

Akshar launched on the App Retailer a number of weeks earlier than this interview. When requested concerning the reception, Permi revealed that the suggestions got here from quarters he hadn’t anticipated. A consumer in Maharashtra wrote to say he was utilizing it to ship WhatsApp messages to his enterprise clients in Marathi. Additional, when Permi posted concerning the app on Reddit, a thread of responses got here in from second-generation Indians in the USA—individuals who grew up talking Indian languages at house, can learn and write them, and have spent their entire lives typing on English keyboards. “They mentioned that that is precisely what they have been searching for,” he recalled.

A language app that’s free for all

The app is free, and Permi says it would stay free. The economics are less complicated than they could seem: as a result of the whole lot runs regionally by way of Core ML, there are not any server prices, no API charges, and no scaling infrastructure. “I’m not having any internet hosting prices or server prices,” he says. “Every part is taken care of by Core ML, working the whole lot regionally.”

When requested if he has plans to deliver the keyboard to Android, Permi mentioned that it’s into account. He works with Swift, an intuitive open-source programming language developed by Apple for constructing apps throughout iOS, MacOS, watchOS, and tvOS. Official assist for growing Android apps utilizing Swift was introduced in October 2025.

Permi shared that he desires to deliver Akshar to Android customers ultimately—not least as a result of a big majority of Indian smartphone customers are on that platform. He’s additionally protecting a detailed watch on AI4Bharat’s analysis output for brand new language fashions. Gondi, an formally recognised Indian language, will not be but supported by IndicXlit. When a mannequin turns into obtainable, he mentioned, he’ll combine it.

When requested what he desires to see change at scale, the keyboard, he mentioned, is a small step towards a broader imaginative and prescient. He sees expertise as a means to assist individuals work together, study, and collaborate with out being restricted by the language they converse.