Inspiration
Americans over 60 reported nearly 4.9 billion dollars in fraud losses in 2024, and AI now lets scammers write flawless phishing emails at scale. Victims rarely know until the money is gone. We wanted protection a family could set up in one click.
How we built it
A family member sends an invite and their relative clicks Continue with Google. On a schedule, Claude Opus 5.5 on Amazon Bedrock investigates every email with a link by checking domain age, Google Safe Browsing and sender mismatches. It then labels scams in Gmail and warns the user and their guardian through SES.
- MCP server: five narrow tools; the agent's only way to touch the world
- AgentCore: Runtime hosts the agent and tools, Gateway routes every call, Identity holds the Gmail tokens, and Policy enforces Cedar rules
- Rest of the stack: DynamoDB, EventBridge, CDK, and a Next.js dashboard on Vercel with no AWS keys
Warnings are permitted only when $R \subseteq {\text{user}, \text{guardian}}$, however the model is prompted.
Challenges
- Our AWS organization blocked the newer Claude endpoint and AgentCore in us-west-2, so we switched to the Converse API and us-east-1.
- Cedar can't read our database, so we generate a policy for each user.
- Gmail has no label-only scope, so the tools and Policy close that gap.
What we learned
Guardrails belong outside the model. We assume it will eventually be fooled and design so that being fooled costs nothing.
Built With
- amazon-bedrock
- amazon-bedrock-agentcore
- amazon-cloudwatch
- amazon-cognito
- amazon-dynamodb
- amazon-eventbridge
- amazon-ses
- amazon-web-services
- aws-cdk
- aws-iam
- cedar
- claude
- gmail-api
- google-oauth
- google-safe-browsing
- model-context-protocol
- next.js
- opentelemetry
- python
- rdap
- react
- strands-agents
- tailwindcss
- typescript
- vercel
