Swalekh
A 22-language Indic typing platform adopted by the Odisha government.
22-language Indic typing platform
Overview
Before AI keyboards were a thing, Swalekh let people type in their own language — 22 of them — in any text field on the web.
A JavaScript plugin that made Indic typing feel native, and ended up running inside a state government's official systems — where it is still live today.
01 — Input
Two ways to type, one keyboard
Two modes — inscript for trained typists, phonetic for everyone else — plus an on-screen keyboard and live transliteration suggestions. Type the way you speak and the right script appears. Next-word predictions sit right on the keyboard, so the next tap is usually already in front of you.
02 — Prediction
No neural net — a trie
The prediction didn't need a model. A trie — a character tree — walks the prefix you've typed and surfaces the most likely next characters and words in microseconds. Simple, fast, and perfect for a plugin that had to run anywhere, on any connection.
03 — Local-first
The device did the remembering
Frequently typed words were cached in local storage and used to re-rank suggestions to each person's own habits. It made the keyboard feel personal — and cut API calls sharply, keeping it fast on slow connections and light on the server.
04 — In production
Inside a state government
The Odisha government wanted Swalekh inside OSWAS — their official workflow system — specifically within its CKEditor rich-text editor. Working with TCS as system integrator, I built a custom CKEditor plugin that brought Indic typing and transliteration into the editor itself. It is still live today.
Under the hood
The transliteration engine started rule-based, then moved to an LSTM model across 11 Indian languages as the data grew. Much of that data came from Lexilor — an internal annotation tool a 24-person team used to build the parallel word-sense corpus the engine learned from.
It shipped before AI keyboards were a thing — and it's still running.
Role
My first product — built end-to-end: the typing engine, the trie prediction, the local-first ranking, and the custom CKEditor plugin that took it into government production with TCS.