Questions people ask.

No. Vakya has had no customers, no clinics and no real callers. Every booking, name and phone number on this site is invented, and the test audio is synthetic speech.

Nothing here is a claim about real clinics. The report card is still empty because Gnani Prisma has not been run yet.

Nothing from Vakya. There is no subscription and no per-minute fee: you run it on your own computer. The code is MIT-licensed and shared on request.

Costs that can appear, and who charges them:

• Speech from Gnani (Prisma to listen, Timbre to speak) is billed by Gnani against your own API key.
• A language model key writes the front desk's replies, and an embeddings key searches its documents. The model provider sets its own limits and prices.
• Phone minutes are billed by your telephony provider (Vobiz), not by Vakya.

• Gnani Prisma recognises what the caller says, and Gnani Timbre speaks, both through Gnani's API with your key.
• A hosted language model decides what to say, and hosted embeddings search the business's documents, with keyword search as the backup. Gnani does the speech, not the reasoning.
• Vobiz carries the phone call.
• There is no second speech model. When Gnani is busy or down, the caller hears a short pre-rendered line and is offered a callback.

First live contact with Gnani was on 5 Nov: Prisma and Timbre work. The realtime stream, cloning and a real phone call have not run yet.

• Hindi, Marathi and English, and callers who mix them in one sentence.
• Prisma and Timbre cover more Indian languages. Vakya has only been tested in these three.
• The front desk answers in the language the caller starts in and follows them if they switch.

On synthetic calls, well. The report card plays the same 24 invented booking calls clean, over an 8 kHz phone line, through street noise and through a phone codec, with and without the business's own names as keyterms, and scores whether the name, day and time came through.

On Gnani Prisma v3.0 (6 Nov 2026) the phone line and the codec cost nothing, and only noise as loud as the speech started to lose a field, which the keyterms won back. Ten-digit phone numbers suffer first in noise.

Even then the audio is synthetic. Repeat it on real recordings before relying on a number. Whatever the table says, the agent reads every number back.

Because phone audio is noisy and a wrong digit in a phone number or a wrong day is a missed patient. Before it books, the agent says the day, date, time and phone number aloud and waits.

It commits only after the caller says yes, and a corrected digit is read back again.

Not through Vakya. Before a voice is cloned, the speaker gives consent: one tick on a plain statement by default, or, in strict mode, a one-time sentence that contains a random code, which is checked against the recording.

The consent record is stored with the voice. One-click withdrawal deletes the voice and its recording, and the withdrawal is written to a signed public log. A clone made with some other tool is outside what Vakya can stop.

Every clip and front desk reply gets a signed receipt, a C2PA content manifest and an “AI-generated” label inside the WAV file. Anyone can drop a clip on the Verify page to check it, and the receipt signature can be checked offline with the public key.

A match is evidence, not proof: it holds only for the exact file, so a clip that has been re-encoded or edited no longer matches its receipt.

Yes. A returning caller is greeted with what they said last time: name, language, last booking, a follow-up note. The profile is kept only to serve their next call, is deleted after a retention period, and the owner can delete a caller by number.

The profile holds the caller's number in full, so treat the store as personal data.

Partly. The front desk can place a call over Vobiz. It speaks first to say it is an AI and stops when the person talks over it. The call logic ran on a live trial line earlier; the Vobiz line is built and tested with fakes, not yet tried on a real call.

Receiving calls on your own number is written but not yet tested live. The missed-call callback, where the front desk rings back a caller who hung up, is built and tested with fakes, not yet tried on a live Vobiz number. Only call people who agreed to be called. There is also a browser web call you can open from a link, to try it without a phone.

It syncs bookings to Google Calendar when a calendar is connected. That sync is written and tested against fakes, not against a real calendar yet.

Without a calendar, bookings are kept in Vakya's own list, and the desk's Bookings page shows them either way. A calendar failure never blocks a booking.

Only the services you give a key to. Your documents, callers and conversations are stored in a folder on your own computer. This site has no trackers or analytics.

Data leaves your machine only through those services: Gnani receives the audio of a call it listens to and the text it speaks, the language model provider receives the text of the turns it handles and the passages of your documents it needs, and Vobiz carries phone calls.

The full list is in the project's status file (shared on request). The short version:

• Gnani Prisma and Timbre first ran live on 5 Nov; the realtime stream, cloning and a real phone call have not run yet.
• The missed-call callback, inbound calls on your own number and real Google Calendar sync are built and tested with fakes, not yet tried on a live line or calendar.
• The report card is empty until Gnani Prisma has been run, and will use synthetic audio, not real callers.
• Caller Theatre has passed 9 of 9 scenarios, on a scripted model only.
• Phone numbers are always read back, so a misheard digit is caught before a booking.
• A receipt matches the exact file it was made for: a re-encoded or edited copy is not recognised.

Vakya builds in what India's rules on synthetic media and personal data are concerned with: consent before cloning, a retention period, one-click withdrawal, an AI-generated label and a signed receipt on everything it makes, and an AI disclosure at the start of every call.

That is a design choice, not a legal opinion. We are a student project, not lawyers, so this is not legal advice. What is lawful for your use, for example calling people, is for you to check.