Loop

Protected call-routing and live scam intervention for families.

Scammers isolate. Loop reconnects.

Loop sits on a call, transcribes both sides in real time, recognizes scam tactics as they unfold, pauses the call and warns the user the instant they're asked for a code or a payment, alerts a trusted family member live, and — once the call ends — plays back a plain-spoken voice recap of what happened.

Two technologies do the heavy lifting:

• Deepgram — real-time, multilingual speech-to-text (and the spoken post-call recap).
• Redis — the vector store for scam-pattern matching, caller fingerprints, and live call state.

For the hackathon, the paid telephony/carrier routing is mocked — but the intelligence pipeline is real.

graph TD
    classDef client fill:#e1f5fe,stroke:#0288d1,stroke-width:2px;
    classDef server fill:#ede7f6,stroke:#5e35b1,stroke-width:2px;
    classDef ext fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px;
    classDef db fill:#fff3e0,stroke:#f57c00,stroke-width:2px;

    subgraph Clients ["App — one app, two roles"]
        UserApp["Protected User"]
        Dashboard["Family Dashboard"]
    end

    subgraph Backend ["Loop Backend (Node + TypeScript)"]
        WS["WebSocket gateway"]
        Pipeline["Live pipeline: rules → risk scoring → protective intervention"]
        Report["Post-call recap (PII-redacted)"]
    end

    Deepgram["Deepgram — nova-3 multilingual STT + TTS"]
    Redis["Redis — vector KNN scam patterns + caller fingerprints + call state"]

    %% Live ingestion
    UserApp -->|dual-track audio| WS
    WS -->|stream audio| Deepgram
    Deepgram -->|live transcript| WS
    WS --> Pipeline

    %% Detection
    Pipeline -->|KNN vector search| Redis
    Redis -->|pattern + caller matches| Pipeline

    %% Intervention
    Pipeline -->|pause + warning| UserApp
    Pipeline -->|live alert| Dashboard

    %% Post-call
    WS -->|call ends| Report
    Report -->|generate voice recap| Deepgram
    Deepgram -->|spoken summary| UserApp
    Report -->|recap + summary| Dashboard

    class UserApp,Dashboard client;
    class WS,Pipeline,Report server;
    class Deepgram ext;
    class Redis db;

Routing is mocked. Intelligence is real.

---

Monorepo layout

loop/
  backend/   Node + TypeScript server (the real intelligence pipeline)
  mobile/    Expo (React Native + TS) iOS app -> TestFlight
  shared/    Types shared between backend and mobile

• The mobile app is the protected call bridge. It captures two audio tracks (caller side + the device mic / user side) and streams them to the backend. One app, two roles: Protected User and Family.
• The backend holds the API keys and runs the pipeline, powered by Deepgram (speech-to-text + voice recap) and Redis Cloud (vector store).

Quick start

See backend/README.md and mobile/README.md.

# backend
cd backend && npm install && cp .env.example .env && npm run dev

# mobile
cd mobile && npm install && npx expo start

Keys

Every integration degrades gracefully when a key is missing, so the app always runs while keys are being provisioned. Fill in backend/.env:

Core (required for the full experience):

• DEEPGRAM_API_KEY — real-time multilingual speech-to-text + the spoken post-call recap
• REDIS_URL — Redis Cloud: vector KNN scam-pattern search, caller fingerprints, call state

Optional (enhance summaries/embeddings; safe to leave blank):

• ANTHROPIC_API_KEY, VOYAGE_API_KEY, TERAC_API_KEY / TERAC_BASE_URL

Multilingual

Scams don't only happen in English. Loop transcribes calls natively across languages using Deepgram's nova-3 model with language=multi, which handles code-switching — a caller mixing Hindi and English ("Hinglish"), or speaking entirely in Spanish, French, or German, all in one stream. Endpointing is tuned to 100 ms so each language shift finalizes promptly. Warnings and the post-call recap are rendered in the conversation's active language, so the protected user is always spoken to in a language they understand.