Check it yourself

Don't take our word for it.

Two things you can try right here, with no microphone and no sign-up, and every result we claim, with the file behind it and the command that reruns it.

Try to fool the guard

Write a call, one line per thing said, or start from one below. Each line goes to the same live guard the app runs during a call, one at a time, in the language you pick. Nothing is spoken and nothing is kept.

Say a number any way you like

Type a UTR, a phone number or an amount the way people say it: digit by digit, in pairs, with “double”, as hundreds and thousands, or mixed with English. This is the same code the agent runs on what it hears; the model never types a digit.

Digits 4021 7788 9910 a 12-digit UTR

What we measured

No real users yet: it was built for this hackathon. So every claim is a back-test on public data or our own break test, labelled as such.

  • 198 of 200

    scam calls in eight Indian languages got a warning, and none of 200 normal calls did

    Keyword rules on the same lines: 16

    Back-test on public data (INDICA, conversations machine-translated into Indian languages)

  • 10 of 39

    legitimate calls got a warning, against 33 of 39 for the same AI model told to warn about scams

    It caught 144 of 200 fresh scam calls; the plain assistant 199, crying wolf on most normal calls

    Back-test on public data (ICFD-31k, synthetic conversations)

  • 24 of 24

    ordinary calls left alone: no warning, no false alarm

    Family, office and payment talk in English, Hindi and Telugu

    Our own break test (we tried to make it fail)

  • 40 of 40

    disguised asks and "stay quiet" tricks still got a warning

    20 disguised asks and 20 attempts to switch it off

    Our own break test (we tried to make it fail)

  • 295 of 300

    spoken 10- and 12-digit numbers read exactly, in noise

    The speech-to-text's own numerals: 79

    Our own break test (synthetic voices, white noise)

  • 207 of 330

    real read sentences in Hindi, Telugu and Kannada with every number exact

    The speech-to-text's own numerals: 114

    Back-test on public data (FLEURS read speech)

  • 211 of 330

    the same sentences through a phone-line codec

    The speech-to-text's own numerals: 115

    Back-test on public data (FLEURS, G.711 phone audio)

  • 540 of 1,107

    read sentences in Tamil, Malayalam, Bengali, Marathi, Gujarati and Punjabi with every number exact

    The speech-to-text's own numerals: 423. Ahead in all six; Bengali nearly level

    Back-test on public data (FLEURS read speech, six more languages)

  • 5 of 5

    scam calls started at the same moment each got the warning

    Each with its own Agora agent; 23.5 to 26.3 s after the press

    Our own test, five calls at once on the live app

How it's built

Everything behind one call: the browser, Agora (RTC, one agent per call with Sarvam and Murf Falcon, the speak API, the second agent for 1930, RTM), our server (the guard, the fast path, the model's 12 tools, the number layer, the screenshot check, the pack) and the services it calls.

Behind the scenes on one call: you, Agora's agent with Sarvam and Murf Falcon, our server's guard, number layer, AI model with 12 tools, screenshot check and pack builder, and the services it calls

See a whole call without talking

The finished complaint, with every step the agent took: sample screen. The 1930 interpreter after the operator gave an acknowledgement number: sample screen. Both use labelled sample data.