Scam Stoper AI

You know how terrifying it is that our loved ones are getting targeted by sophisticated AI scams and losing their life savings? We're building Scam Stopper AI to fix that.

Comment

Scam Stopper AI — voice-first protection against AI-driven scams

A conversational, voice-driven assistant that helps older adults recognize scam patterns in real time and follow safe next steps.

Inspiration

AI scams are getting unnervingly believable—especially over the phone. We built Scam Stopper AI to give families a calm, voice-first safety layer that can detect common scam tactics (urgency, secrecy, payment redirection) and guide someone through what to do next.

What it does

Scam Stopper AI lets a user speak about a suspicious call/text/email and then:

  • Asks quick clarifying questions (who contacted you, what they asked for, payment method, time pressure)
  • Produces a Risk Level: LOW / MEDIUM / HIGH
  • Explains the signals that drove the decision (e.g., “urgent deadline”, “OTP request”, “gift cards”)
  • Generates a step-by-step action plan
  • Generates a safe reply script the user can read verbatim
  • (Optional) Creates a family summary to share with a trusted contact

Demo

How we built it

Architecture (high level)

Layer Tech Purpose
Frontend React Simple UI + “Talk now” flow + risk meter + checklist
Voice agent ElevenLabs Agents Natural voice conversation + tool calling
Backend Google Cloud Run HTTPS endpoints (“tools”) used by the agent
AI reasoning Vertex AI (Gemini) Scam classification + signals + actions returned as strict JSON
Storage (optional) Firestore / Cloud Storage Incident history + uploaded evidence

Workflow

  1. User speaks → ElevenLabs Agent handles the conversation
  2. Agent calls our Cloud Run tool endpoint with the transcript
  3. Cloud Run calls Gemini on Vertex AI and returns structured JSON
  4. Agent reads results aloud + UI renders them as a checklist

Challenges we ran into

  • Reliable structured outputs: We needed consistent JSON so the UI wouldn’t break.
  • Safety guardrails: Ensuring the assistant never asks for sensitive data (passwords/OTP/SSN) and never suggests unsafe actions.
  • Latency: Voice UX must feel real-time, so we optimized model choice + minimized tool calls.

What we learned

  • How to design voice agents that feel natural while keeping the backend deterministic.
  • How important guardrails and “verify via known channels” defaults are for high-stakes scenarios.

What’s next

  • Voicemail + SMS ingestion (upload/forward and analyze)
  • Multilingual support and regional scam patterns
  • A family dashboard for incident history and prevention coaching
  • Reporting templates (FTC/IC3) and optional observability

Built With

  • ElevenLabs Agents
  • Google Cloud Run
  • Vertex AI (Gemini)
  • React
  • (Optional) Firestore, Cloud Storage

Built With

Share this project:

Updates