# Samvaad

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Yash Mittal](https://cerebralvalley.ai/u/yashmittal)
- **GitHub:** https://github.com/yassmittal/samvaad
- **Demo video:** https://youtu.be/jQXBUj_tNRU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2486

Samvaad (Hindi for "conversation") is a phone call that translates itself. Two people who don't share a language each speak their own and hear the other in theirs, about 2 to 3 seconds behind.

The problem: a lot of everyday business in India happens across a language gap, such as an exporter and a foreign buyer, a shopkeeper and a tourist, or a migrant worker and a landlord. India has 22 scheduled languages and most people mix two of them in the same sentence (Hinglish). Phrase-by-phrase translators like Google Translate's conversation mode take each sentence on its own. They forget names, product terms and what was said a minute ago, and they handle code-switching badly.

How it works: you start a call, pick your language and send the link on WhatsApp. The other person taps Answer in their browser, with no app and no sign-up. When you speak, Muse Voice Transcribe streams your words and knows when your sentence ended. Muse Spark translates it knowing the setting, a glossary and the last eight sentences, and the other phone says it in a natural voice. Ticks like a chat app show "Translated" and then "Heard", and the screen shows the delay for every sentence. When the call ends, Muse Spark writes a recap in each person's language. Table mode lets two people share one phone on a table.

It is hosted and works today: the web app is at https://samvaad.yashmittal.xyz and the API runs on EC2. The demo video is an unedited call between two Android phones. Measured on the hosted app, it takes 2.3 to 2.6 s from the end of a sentence to the other phone speaking.

Everything was built during the hackathon, solo, starting from an empty repo on 29 Oct: the call UI, WebSocket rooms with reconnects, streaming to both Meta models, context-aware translation, the recap, the natural voice, WebRTC for the other person's real voice, and hosting. It has 289 tests, and every design decision is recorded in docs/decisions.md.

## More from Meta Global AI Developer Hackathon

- [Trail Sentinel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2483)
- [Guiding Hand](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2484)
- [Nomi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2485)
- [MuseNative](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2487)
- [Sravan Kumar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2488)
- [OpenAgents](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2489)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2486. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt