Skip to content

Kavach

Kavach — real-time call intelligence that hears the manipulation, not just the words.

Core trick

Live 16 kHz call audio streams to Gemini, where schema-validated tool calls for risk signals pass a two-hit state machine before any warning.

Tags

Description

Kavach — real-time call intelligence that hears the manipulation, not just the words. India loses over ₹22,000 crore a year to scam calls that follow predictable scripts: hook → authority → isolation → urgency → extraction. Kavach listens to a live call and intervenes before the money moves. Real-time & multimodal: the phone streams live 16 kHz audio to Gemini 3.1 Flash Live, which listens continuously — reading vocal urgency, catching the caller mid-sentence, and speaking a calm counter-warning into the call the moment the script turns coercive. It isn't turn-based; the caller can interrupt it and it keeps listening. Each risk signal is a schema-validated tool call gated by a local two-hit state machine, so the model senses and reasons but never acts alone. On-device Gemma 4: when the network drops — exactly when a first-time victim is most exposed — a Gemma 4 E2B agent runs natively on the phone via LiteRT. It's a full sense → decide → act → check loop that holds state across the call, recovers from malformed model output, and defers to a human at clear boundaries. It keeps working even with no weights loaded. Beyond detection: Kavach names the manipulation tactic and explains it in one plain sentence, scores caller trust live, can flip to a honeypot persona that wastes the scammer's time, and writes a post-call forensic debrief. Every event lands in a privacy-preserving, tamper-evident local ledger — transcript hashes, never raw audio. Verified with real audio on gemini-3.1-flash-live-preview; a 16-case regression scores 1.00 precision and recall. Dual-brain: Gemini when connected, Gemma on-device when not.
krisha-oswal/kavach★ 00 forksKotlinCreated Jul 11, 2026 · last push Jul 11, 2026

README excerpt

# Kavach — real-time call intelligence

Scam calls aren't improvisation, they're **scripts with acts**: hook → authority →
isolation → urgency → extraction. Kavach listens to a live call and, powered by
Gemini, does four things a keyword detector can't:

- **Understands & explains** — Gemini reads the transcript and says, in one plain
  sentence, *how* you're being manipulated ("pretending to be a CBI officer to scare
  you into believing you're under investigation").
- **Scores trust, live** — a caller starts at 100 and loses trust as tactics stack.
  The gauge falls in real time and the room sees *why*.
- **Acts** — warns you in your language, or flips to **Honeypot mode**: Kavach picks up
  as a rambling elderly persona (dynamic, Gemini-driven) and wastes the scammer's time.
- **Learns & debriefs** — a **Scam Radar** shows what's circulating; a post-call
  **Forensic Reveal** (Gemini) explains what happened and generates a shareable
  warning card.

Dual-brain: **Gemini when connected, on-device Gemma/rules when not** — pull the
network and it keeps working. (PS1 + Gemma Special Prize.)

## The Google models it uses (all live on the hackathon key)
| Job | Model |
|-----|-------|
| Live analyst + explanations | `gemini-3.5-flash` (thinking off → ~1.5s, structured JSON) |
| Dynamic honeypot persona | `gemini-3.5-flash` |
| Post-call forensic debrief | `gemini-3.5-flash` |
| Shareable warning card | `gemini-3.1-flash-lite-image` (NB2 Lite) |
| Live audio (optional) | `gemini-

Team