Inspiration

Americans reported losing over $12.5 billion to fraud in 2024, and the worst scams don't arrive as email. They come as a text about an unpaid toll, a call from "Coinbase security", or a panicked "Grandma, it's me, I'm in jail." We wanted protection that works inside that moment, in the apps people already use: Messages and the Phone app. No new app, nothing to install, and simple enough for our grandparents. And because T-Mobile sees scams across the whole network, the protection should get smarter for everyone each time someone reports one.

What it does

Bouncer is an AI scam bodyguard that you save in your contacts.

  1. Show it anything suspicious: Describe something questionable, send a screenshot, or paste a DM or email, and Bouncer will reply with a plain-English verdict, multiple red flags, and safe next steps.
  2. Merge it into a sketchy call: Bouncer joins your phone calls discretely and checks every part of the conversation for scam signs. It texts you warnings about the call while it is happening, and gives you a short summary afterwards.
  3. Family Safety: Set up a secure family circle by text with a shared safe word. If someone messages "Mom is calling me crying, saying she was kidnapped", Bouncer will then call her real phone to check for the safe word, and make sure she's okay before alerting the rest of the family.

How we built it

Server: Node.js 22 and Typescript on Hono with PostgreSQL, deployed on Railway Messaging and voice: Photon 2.0 for both iMessage and voice, with the voice line connected to an ElevenLabs Agent over SIP. AI Models: TypeSafe Jev makes the scam decisions by answering targeted questions (impersonation, urgency, requesting 2FA codes, etc.), and responds with a calibrated risk level. DeepSeek V4.1 Flash reads screenshots, text and voice messages, and responds to the user without determining whether something is a scam.

Challenges we ran into

  • Scammers can interact with Bouncer: On a merged call, anybody on the line can call for Bouncer's help, which means that scammers could potentially manipulate and persuade the agent. Because of this, we decided that nothing important should rely on voice alone. If you need to change your family's safe word or obtain personal details, it must be done over iMessage.
  • Classification: The classification that Jev provides can incorrectly flag a large percentage of situations if the question isn't asked correctly or the answer isn't framed properly. For example, when flagging impersonation attempts, the classifier flagged every message claiming to be from a legitimate company as impersonation, whether it was correct or incorrect. Because of this, we had to add a rule that the 'Likely scam' classification had to have at least one other red flag on top of impersonation.

Accomplishments that we're proud of

  • Bouncer calls your family and makes sure they're safe, which is extremely impactful as AI voice cloning scams rise.
  • 98% of the scam texts used in our datasets were caught by Bouncer on 1,000+ messages we didn't use for tuning.
  • No normal calls were given a false positive, across both our datasets and personal testing.

What we learned

  • Once calibrated, a classification model like Jev can provide intelligence comparable to many state-of-the-art models while staying deterministic.
  • Every change and fix made during tuning had to be measured with seemingly obvious changes making our evaluations worse.

What's next for Bouncer

  • Scam Shield at the carrier level: With T-Mobile's network behind it, Bouncer could catch scam texts as they arrive, spot scams spreading by ZIP code, and warn people before the scam reaches them. Nobody would have to forward anything. We built a mock-up of this carrier console using simulated data.
  • More channels: SMS and RCS for Android, plus WhatsApp, Telegram, Facebook Messenger and email. Only the messaging layer is tied to iMessage, so the scam checks carry over as they are.
  • Privacy and consent: In states where everyone on a call must agree to being listened to, Bouncer would start the call with "This call is being screened for scams." We'd also need vendor contracts that forbid training on call data, a T-Mobile privacy review, and STOP and DELETE MY DATA commands.
  • Sign-up at scale: Turn Bouncer on from your T-Mobile account in the T-Life app or a store.

Built With

Share this project:

Updates

Submission history