Teaching AI to Speak 15 Indian Languages

Every feature we ship needs its words translated into ten or more languages. Not just video copy: the UI copy in Figma too. Here's how that used to work: write it in English, send it to an outside agency, wait, check what comes back, repeat every time we launch something. It works, but every launch waits on the agency's clock. And it's hard to keep the tone the same across that many hands.

For one of ShareChat's pro features, the Creator Earning Program, I pitched a different way: build a small AI agent in Figma, and give it a clear set of rules to follow. What words to keep as-is, what to actually translate, how the word "earn" changes meaning depending on the sentence, what tone fits the brand in each language. Write the rules once. Let the AI follow them every time. The copy team wrote the rules, and the copy team built it.

"Why should we do this?" That was the first thing people asked. Not a no, a real question about whether the setup was worth it.

I didn't have a clean answer beyond believing it was right. This would let the whole design team move faster, and once it worked, it wouldn't just help one feature, it would become how we do all of it. That's a harder thing to sell than a number. I stuck with it anyway.

• 01Wrote the rules: the agent decides what to translate and what to keep as an English word, written in the local script. "Start Earning" gets truly translated. "Views" stays "views," just written in Tamil letters. And "earn views" becomes "get views," because that's how people actually say it. Keep the tone short and modern.
• 02First try, Malayalam. A native speaker checked it, told us what was wrong, we fixed it, tried again.
• 03Second try was about 98% right. Sentences read naturally, only a few tricky words were off. We put it into real use: the AI does the work, a writer just checks it.

That part worked well. Then we tried to do the same thing for more languages, and it broke.

"The Figma method didn't work past the first language. More languages meant more screens, more reviewers, and more to coordinate than we could handle."

We tried something lighter instead: take the same rules, upload them to Gemini along with the Malayalam version as an example, then translate into Marathi and Punjabi using a spreadsheet instead of Figma. Easier to check with more people at once. But the result was only about 85% right, worse than the 98% we got on Malayalam.

The problem wasn't the tool. It was which language we used as the example. Malayalam got better when we used Tamil as the example instead of English, because Tamil and Malayalam are close cousins, from the same language family. But for Marathi and Punjabi, we still used Malayalam as the example, and Malayalam comes from a different family than both of those. Wrong example language, lower quality. The fix seems obvious now, but wasn't obvious before we tried: pick your example language from the same family as the one you're translating into.

The limits were real at first. Every language needed a native speaker to check it. Every new feature needed its own rules, and the example language mattered more than we expected. We worked through them instead of giving up on the method.

"One language didn't just prove the idea. It proved the principle: pick your example from the same script family, and the next language follows."

Where it stands today: the copy team runs the Figma agent with skills as the way translation gets done, and accuracy sits around 95% across languages. The lesson that unlocked it was the script family. Malayalam didn't just work on its own, it showed that every language works when its example comes from the same family. We're now rolling the same method across all videos and translations, and exploring the UI and UX copy in Figma next. The agency wait is gone from the launch path.