ScamSaver

Think before you act, We'll think with you.

Comment

Inspiration

We get suspicious calls and texts constantly, from unknown numbers to "Congratulations! You have won $x " messages. One of us got one this morning before the hackathon even started. It got us thinking about how there's no quick way to know in the moment whether something is real or not. Americans lost over $12.5 billion to scams in 2025 alone, and most people just have to guess. We wanted to build something that takes the guesswork out of it.

What it does

Paste a suspicious text, upload a screenshot, upload a voicemail, or press one button during a live call and ScamSaver analyzes it in real time. It returns a scam probability score, highlights suspicious phrases, and tells you in plain language what to do next.

How we built it

OpenNote for brainstorming and researching scam details which were used as input.

  • Frontend: Next.js, Tesseract.js for screenshot OCR, Web Speech API for live transcription, Recharts for real-time scam breakdown visualization
  • Backend: Next.js, Featherless.ai with Llama 3.1 8B for scam detection, Deepgram for audio transcription
  • Deployment: Vercel

One of us handled all the backend and AI integration, the other built the entire frontend. GitHub branches kept us from stepping on each other. We later worked on one laptop for both backend and frontend due to one of us having device implications.

Challenges we ran into

ElevenLabs went down mid-hackathon and took out audio transcription. We swapped in Deepgram and within 30 minutes and didn't lose any features. However, for live transcription, getting the live call listener to feel genuinely real-time without chunks being too short to transcribe was also way trickier than expected. We spent almost 2 hours trying to understand why the live transcription was not working and ended up using Web Speech API for the live transcription instead of Deepgram.

Accomplishments that we're proud of

All the features we created ESPECIALLY The live call feature. Press a button during a suspicious phone call and AI tells you in real time whether you're being scammed. That's what we came in wanting to build and we shipped it.

What we learned

Always have fallbacks for third party APIs. ElevenLabs taught us that the hard way but it made the whole system more resilient. We also got a lot better at parallel development, although one of us made a bad commit and it messed up a lot of the workflow.

What's next for ScamSaver

Phone number lookup before you even pick up, a browser extension for suspicious emails, multilingual support for non-English speakers, and a dashboard to track and log scam attempts over time.

Built With

  • deepgram
  • featherless.ai
  • llama
  • next.js
  • node.js
  • react
  • recharts
  • webspeech
Share this project:

Updates