Gnani AI · Great Indian AI Internship Challenge 2026

A second listener for scam calls.

Kavach catches scam calls in Telugu and Hindi, warns the listener, alerts the family and seals the evidence on Bitcoin. Built entirely on Gnani AI: Prisma, Evon and Timbre.

₹22,845 crorelost to cyber fraud in India in 2024 (MHA, Lok Sabha, July 2025)
₹62 crore a dayThe typical attack is a phone call asking for an OTP
8 agentsThe LLM argues both sides. Code decides.

See it work

The full demo is one real run of the app, narrated: a Telugu scam call, the verdict, the warning, the family alert, the 1930 complaint, the seal, a genuine call left alone, and the family asking.

Full demo · 2:21
Trailer · 0:60

Try the real screen

This is Kavach's own screen, replaying three saved Telugu calls that it analysed with Gnani Prisma, Evon and Timbre. Press Analyse and the call plays, with English subtitles:

  • the caller's words that prove the scam, underlined
  • what the Prosecutor, Defender and Judge each decided
  • the Telugu warning and the family's voice note
  • the 1930 complaint, and the seal checked against Bitcoin
  • a genuine bank call, left alone
Open the replay →

Live analysis needs our Gnani key and Evon on a GPU, so it runs on our machine, not here.

Eight agents, one job

During the call, agents 1–7 run in a fixed order. Evon argues; the verdict is decided by code, never by the LLM.

1ListenerGnani Prisma

Transcribes the call and separates the caller from the listener.

2ProsecutorGnani Evon

Lists the scam tactics, quoting the caller.

3DefenderGnani Evon

Lists signs the call is genuine.

4JudgeCode

Counts a claim only if the caller's own words prove it, and it isn't negated.

5GuardianEvon + Timbre

Warns the listener in Telugu or Hindi, and alerts the family with a voice note.

6ReporterGnani Evon

Drafts the complaint for the 1930 cybercrime helpline.

7NotarySHA-256 · Bitcoin

Seals the evidence in a hash chain, timestamped with OpenTimestamps.

8Family assistantEvon over MCP

After the call, answers the family's questions from the sealed evidence.

Why it can be trusted

  • The LLM argues; code decides. An LLM judge flagged every genuine call. With the prosecutor/defender split and a code Judge, no genuine test call has been flagged since.
  • The caller's words never reach the listener. Warnings and alerts are built only from verified tactics.
  • Only a 32-byte hash leaves the device. A one-byte edit fails verification against the Bitcoin block.

Results so far (synthetic calls)

22/22Genuine Telugu–English calls left alone
16/22Scams caught on the same set
3/3Telugu calls right end to end with real Evon
9–12 sTime to a Telugu verdict
120Automated tests passing

100% Gnani, sovereign AI

Prisma v2.5Speech-to-text in Telugu, Hindi, English and code-mixed speech; separates caller and listener
Evon v3.3Open-weight LLM, self-hosted
TimbreSpeaks the warning and the family's voice note, in Telugu and Hindi

Limits and next steps

  • Calls are analysed after recording. Next: streaming Prisma, for warnings during the call.
  • Family alerts are built, voiced and queued; no SMS/WhatsApp gateway is connected yet.
  • All test calls are synthetic. Next: more Indian languages and a pilot with real families.