FraudGuard.AI was inspired by the rapid rise of AI-powered voice scams and the lack of accessible training for everyday people.
While corporate companies have employees go through training by using simulated phishing- the elderly, students who are just now entering the adult world, and regular people with busy lives and many things to worry about have no access to similar tools. We wanted to make realistic scam training available to anyone with just a phone.
FraudGuard.AI calls the user’s phone simulating a realistic AI scam call and then assigns a vulnerability score based on how they responded. During the call, the platform streams a live transcript and detects risky disclosures and behaviors. Afterward, users receive a score, a color-coded transcript, and specific guidance on how to respond safely in real situations.
We built a React frontend and an Express.js backend connected through Twilio to place calls and ElevenLabs to generate the scammer’s voice. Transcripts stream in real time via WebSockets and are analyzed using a Groq-hosted LLM to detect risk signals. A deterministic scoring engine converts those signals into a final score, which is stored anonymously using Firebase with a random session ID.
One major challenge was reliably scoring open-ended human conversation without false detections. We addressed this by combining LLM extraction with strict backend scoring rules. Integrating real-time voice AI, telephony, and live transcript streaming also required careful coordination to maintain low latency and a smooth user experience.
We’re proud that we built a complete end-to-end system that can call a real phone, simulate a convincing scam, and instantly generate a personalized risk assessment. The platform delivers a memorable training experience and provides clear, actionable feedback instead of generic advice. It also works instantly without requiring accounts or setup.
We learned how to integrate real-time voice AI, telephony infrastructure, and LLM analysis into a single workflow. We also learned that deterministic backend logic is essential for making AI-driven scoring consistent and explainable. The project gave us a deeper understanding of how social engineering exploits human behavior.
Next, we plan to add more scam scenarios, adjustable difficulty levels, and multilingual support. We also want to allow optional accounts so users can track progress over time while preserving privacy. Our long-term goal is to make scam awareness training accessible and routine for anyone.
Built With
- css-frontend:-react-backend:-node.js
- elevenlabs-api
- elevenlabs-conversational-ai-ai-/-nlp:-groq-(llama-based-models)-for-transcript-analysis-and-risk-extraction-realtime:-websockets-for-live-transcript-streaming-database:-firebase-firestore-cloud-/-hosting:-firebase-(database-and-backend-services)
- express.js-voice-&-telephony:-twilio-programmable-voice
- groq
- html
- languages:-javascript
- typescript
- vercel-(frontend-deployment)-apis:-twilio-api
Log in or sign up for Devpost to join the conversation.