Scroll down to see how it works
How it works
Listen in on one call. You and a friend speak in two languages; Aaytu hears who said what, checks your saved places and contacts, waits for your yes and PIN, and gets it done.
You: Okay, bye.
Aaytu
Okay, bye!
Behind the call
- 1
Prisma hears
English speech
- 2
Aaytu decides
End the call
- 3
Safety check
Spoken yes + PIN match
- 4
Timbre speaks
Replying in your language
Done on this call
- Done
Swiggy orderSimulated
Bakingo · KitKat cake, dosa · ₹149
- Done
Email sent
To Manager · “Running late to the meeting”
- Done
Call ended by Aaytu
Everything is in your call log
1. Link once
Sign in, verify your number, add contacts and the places you order from.
2. Call
Any phone. The assistant picks up, listens and acts.
3. See it done
Every call and task in your app, with the transcript.
Models
Built on Gnani's models and APIs for the Great Indian AI Internship Challenge.
Hears
Prisma v2.5
Gnani's speech-to-text, built for 8 kHz phone audio. We run one Prisma stream per language on the same call audio — English, Kannada and Hindi — and pick the best transcript for each sentence, so a friend can join in another language mid-call.
Speaks
Timbre v2.5
Gnani's text-to-speech, streamed over WebSocket so the reply starts in about a second and the caller can interrupt it. Replies come back in the language the speaker used.
Understands and decides
Evon v3.3
Gnani's 30B open-weights language model, served on our own GPU. It understands the request, translates, and picks one tool per turn.