AEGIS Scam Guard

A WhatsApp-native AI agent that checks suspicious messages, links, images, and voice notes for scam risk using Gemini and Google Cloud.

Comment

Inspiration

Scam messages move faster than help. A person can receive a fake bank alert, a phishing link, an impersonation message, or a suspicious image on WhatsApp and be pressured to act before they can verify it.

AEGIS Scam Guard was built for that exact moment: when someone needs a fast, clear, trustworthy second opinion inside the app they already use.

What it does

AEGIS Scam Guard is a WhatsApp-native AI safety agent. A user forwards a suspicious message, link, image, or voice note to the bot, and AEGIS analyzes it for scam and phishing risk.

The agent replies in plain language with:

  • Whether the content appears safe, suspicious, or scam-like
  • The specific manipulation patterns it detected
  • Why the message may be dangerous
  • Suggested next steps the user can take
  • Optional follow-up behavior for continued protection

Unlike a normal chatbot, AEGIS runs as an asynchronous cloud agent. It receives WhatsApp webhooks, publishes reports through Pub/Sub, reasons with Gemini, stores anonymized threat memory in Firestore, and can continue monitoring cases in the background.

How we built it

AEGIS is deployed on Google Cloud as a production-style, event-driven agent system.

The architecture includes:

  • WhatsApp Business Cloud API for real user interaction
  • Cloud Run ingestion service for webhook verification and message normalization
  • Pub/Sub for asynchronous report processing
  • Cloud Run agent orchestrator running a Google ADK agent
  • Gemini on Vertex AI for scam-risk reasoning
  • Firestore for case state, threat memory, and follow-up records
  • Cloud KMS and Secret Manager for secure contact handling and credentials
  • Cloud Scheduler and a follow-up worker for background re-engagement
  • Terraform for reproducible infrastructure
  • A React/Vite web page for public project presentation

The system is intentionally decoupled: ingestion does not call Gemini directly, the agent does not receive raw webhook traffic, and sensitive user identifiers are hashed before storage.

What makes it agentic

AEGIS is not just a prompt wrapper. The agent follows a structured workflow:

  1. Analyze the forwarded content
  2. Detect manipulation patterns such as urgency, impersonation, phishing, or financial pressure
  3. Check sender and threat reputation
  4. Cross-reference anonymized reports
  5. Draft a protective response
  6. Escalate, close, or schedule follow-up action

This lets the system move from passive chat to active protection.

Challenges we ran into

The hardest parts were production reliability and privacy.

We had to handle real WhatsApp webhook signatures, media downloads, Pub/Sub delivery, Cloud Run IAM, encrypted contact resolution, and Vertex AI quota limits. During testing, we also found and fixed issues around Firestore timestamp serialization, partial document updates, Pub/Sub authentication, and model quota failures.

For example, image messages were reaching the backend correctly, but a Vertex AI 429 RESOURCE_EXHAUSTED error could stop the user from receiving a reply. We added a fallback path so the bot now sends a visible WhatsApp message when the model is temporarily rate-limited.

Accomplishments

We are proud that AEGIS is deployed as a real working system, not only a mock demo.

The project includes:

  • A live WhatsApp bot
  • A live Google Cloud-hosted project website
  • Real Cloud Run services
  • Pub/Sub-backed asynchronous processing
  • Firestore-backed memory
  • KMS-based contact protection
  • CI tests and Docker builds
  • Production screenshots showing Cloud Run traffic, logs, and WhatsApp replies

What we learned

We learned that building an agent for real users is mostly about the system around the model: queues, retries, privacy boundaries, logging, quota handling, secure credentials, and clear user-facing failure modes.

A good AI agent is not only intelligent. It must be dependable when the network, APIs, and users are unpredictable.

What's next

Next, AEGIS can expand into:

  • Better multimodal image and voice-note scam analysis
  • Regional scam pattern intelligence
  • Family-safe opt-in alerts
  • Browser and SMS integrations
  • Public scam trend dashboards
  • Stronger abuse prevention and rate limiting
  • Verified partnerships with consumer protection groups

AEGIS Scam Guard shows how Gemini and Google Cloud can turn everyday messaging into a real-time protection layer for people at risk of digital fraud.

Built With

Share this project:

Updates

Submission history