NFT AtomsAPPLIED AI & PRODUCT ENGINEERING
Explore the voice demoDiscuss your use case
Menu
Voice AI · Live demo

Dhwani, a phone agentyou can call today.

Dhwani holds a real phone conversation and switches between Hindi and English mid-sentence, the way people in India actually talk. This page is a live demo of how it listens, answers and stays within its rules.

Launch the demo Request access
  • 770 ms p90 time to first token, on our own logged calls
  • Hindi and English in the same sentence
  • Interruptible mid-sentence, like a person
  • Guardrails enforced in code, not in the prompt
Dhwani · voice pipeline

01 / 05 · Call

The phone rings.

A real call arrives over the network. The agent picks up and is ready to listen in well under a second.

This is a demo, not a product

Fictional scenario

Every person, account and rupee amount you hear is fictional, written for this demonstration. Dhwani here is not a collections service NFT Atoms operates or sells, there is no real borrower or lender behind it, and nothing you say is used for anything. Please don’t share real account or financial details. Needs a microphone and works best in Chrome. The link is access-code protected: request access if you don’t have one.

The numbers

Measured onour own calls.

These come from our own test calls, after trying several model combinations on real logged conversations. They are not a guarantee.

Time to first token, p90
1,761 ms

Down from 1,761 ms before the bake-off.

Faster than before
0%

The same calls, the same scenario, a different model mix.

Typical call length
1–5 min

Long enough to be a real conversation, not a prompt.

What you’ll hear

What we built intothe call.

01

Answers before you finish the sentence

  • Speaks while it’s still writing: the model’s output streams straight into speech, so the first word arrives before the sentence is finished.
  • Two speech-recognition providers race on every turn. Whichever returns first wins, and the other is the fallback.
  • Transcribes in the pause, not after it. Speculative recognition starts on the natural gaps in speech.
02

Turn-taking that doesn’t feel scripted

  • Backchannels (“hmm”, “ji”) while you’re still talking, the way a listener does.
  • Can be interrupted mid-sentence: it stops immediately, like a person would.
  • Waits exactly as long as the question deserves: a shorter pause after a yes-or-no, a longer one after a number.
  • If you pause mid-thought and keep talking, it discards its half-formed reply and listens again.
03

Guardrails as a feature, not an afterthought

  • A validator blocks any number the agent didn’t get from the account record, so it cannot invent an amount or a date.
  • It cannot be talked into claiming to be someone or something it is not.
  • Identity has to be genuinely confirmed before any account detail is discussed.
  • Third-party non-disclosure and distress handling are enforced by code, not by prompt.
Outcomes

Every call ends witha recorded outcome.

Each call finishes as one of six outcomes, logged with the latency of every turn and the compliance checks that ran. That log is how we audit it, and how we make the next version better.

  • Promise to pay
  • Callback requested
  • Dispute
  • Wrong number
  • Distress detected
  • Handoff to a human
Before you call

A few things to know.

Is this a real debt-collection service?

No. This page is a technology demonstration only. The borrower, the lender, the account and the amount owed are all fictional, written for this scenario. NFT Atoms doesn’t operate or sell a collections service, and no real accounts, debts or consumers are involved.

Should I use real information when I call?

No. There’s nothing real on the other end to give it to. Treat the call as a scripted, fictional scenario built to show how the agent handles a hard conversation, not a real transaction. Please don’t share real account or financial details.

Why does it need a microphone, and why Chrome?

The demo is a real-time spoken conversation running in your browser, so it needs microphone access to hear you. It’s currently verified to work best in Chrome on desktop.

Why is the demo access-code protected?

It’s a private link while we finish rolling it out. Request access and we’ll send you a code.

Can it actually mix Hindi and English mid-sentence, or does it just switch modes?

It code-switches the way bilingual speakers naturally do (Hindi and English inside the same sentence), rather than detecting one language up front and locking into it for the whole call.

Get in touch

Tell us aboutthe work.

A workflow that takes too long, a model that keeps missing, a product that needs its next version. Write to us and we reply within a working day.

Discuss your use case hello@nftatoms.com