Building AI Agents for Hindi, Marathi and Hinglish
Most Indian customers do not write pure Hindi or pure English. They write "order kab tak aayega?" An agent built only for English misses a large share of real conversations.
- Published
Text agents
- Detect language per message and reply in the customer’s language and script — Devanagari or romanised.
- Handle code-mixing: modern LLMs understand Hinglish well, but test on your real messages.
- Keep retrieval language-aware: a Hindi question should still find an English policy document.
- Budget for higher token counts in Indian scripts.
Voice agents
Speech recognition quality varies widely across Indian languages and accents. We benchmark several speech-to-text providers on recordings that resemble your real callers — including background noise and phone-line quality — before choosing. Names, addresses and numbers need special handling and confirmation.
Evals per language
| Language | What we test |
|---|---|
| English | Baseline accuracy |
| Hindi (Devanagari) | Understanding and reply quality |
| Hinglish (romanised) | Code-mixed understanding |
| Marathi / others | Understanding, script, tone |
Accuracy often differs by language. Measure each separately and set targets per language rather than one blended number.
Tone and context
Politeness norms differ: "aap" versus "tum", honorifics, formality with elders. We write style guidance per language and review sample conversations with native speakers.
Frequently asked questions
Which Indian languages can you support?
Major Indian languages for text; for voice, we confirm quality per language during a benchmarking step.
Does multilingual support cost more?
Somewhat — mainly eval coverage per language and slightly higher token use.
Can one agent handle several languages?
Yes — one agent with language-aware prompts and evals is the norm.
