Building multilingual AI apps for India

India has 22 scheduled languages and a wide range of digital literacy. If an AI product only works well in English text, it leaves out a large share of the people who could benefit most from it. Three of my projects tackled this from different angles: MediSage, AgriFinAI and the Saathi Smart Assistant.

Here's what I learned building them.

1. Let users choose the language explicitly

MediSage is the simplest of the three: a Streamlit app where you choose a language, enter a medicine or medical condition, and get a generated explanation from Google's Gemini model.

The explicit language selector matters more than it looks. Auto-detecting language from a short input — a single medicine name, for example — is unreliable. Asking up front is simple, predictable and respectful of the user's preference.

2. Separate UI translation from AI translation

AgriFinAI combines a Gemini-powered assistant with a loan marketplace, crop and weather advisory and soil analysis. It supports languages including Hindi, Marathi, Tamil, Kannada and Gujarati.

There are two different translation problems in an app like this:

• Interface text — buttons, labels, navigation. This should be deterministic, so AgriFinAI uses i18next with translation files.
• Generated answers — these come from the model, so the model is instructed to respond in the user's selected language.

Mixing these up leads to an app where the chatbot speaks Tamil but the "Send" button is in English, or where interface strings change every time they're generated.

3. Voice-first for low digital literacy

Saathi was designed around customers in Tier 3 and Tier 4 regions. For many users, typing in any language is the barrier — so the interface is voice-first, with text as a fallback.

Voice-first changes the design:

• Responses need to be short enough to listen to.
• Every flow — onboarding, payment questions, grievances — must work without reading.
• Offline behavior matters; Saathi includes a service worker for offline support.

4. Keep answers short and structured

In AgriFinAI, the assistant is instructed to answer in roughly 100–150 words and in a consistent format. Long, meandering answers are hard to read on a phone and even harder to listen to. A tight word budget forces the model to prioritize.

5. Make the AI explain itself

AgriFinAI's assistant is designed to cite sources and explain its reasoning. When the advice affects someone's crop or loan decision, "trust me" isn't good enough. Grounding answers in public datasets and showing why a recommendation was made builds the trust needed for people to act on it.

6. Be clear about limits in sensitive domains

Health and finance answers can cause real harm if they're wrong. MediSage is an informational experiment, not medical advice, and a product like it should say so plainly. The same applies to financial guidance.

The common thread

Multilingual AI isn't just "call the model and ask for Hindi". It's choosing what to translate deterministically, what to generate, how long answers should be, how users provide input, and how much the system should explain itself. Getting those choices right is what makes an AI product genuinely usable for more people.