A second opinion before you pay, through your Meta AI glasses.
Look at a letter, a text, a webshop or listen to a call. In about two seconds you hear whether it shows scam signs, and why.
▶ Watch the video · Try it in the browser · iOS app · Server
Scams don't wait in your spam folder. They come on paper, on screens and on the phone, and they want more than a payment: a bank code, a password, your details, or remote access to your computer. Most arrive where no filter sees them.
| €72.5M | lost to scams in Finland in 2025, up from €62.9M in 2024 |
| €38.4M | of that taken by fake "safe account" calls |
| 57% | of adults worldwide met a scam last year; 23% lost money |
Sources: Finance Finland (bank data), 20 Feb 2026; Global Anti-Scam Alliance, Global State of Scams 2025. Details in presentation/LEDGER.md.
Everyone thinks they'd never fall for one. That's the point.
- Just look. Tap check or press the glasses' capture button. One photo, an answer in about 2 seconds, spoken through the glasses.
- Always on, if you want. A quiet check every 15 seconds; it only speaks when something looks wrong.
- It tells you why. One verdict, what to do, one plain reason and the signs it noticed. "Uses the name 'ray-ban.com' but ray-band.com is not its official website."
- It checks what scam words miss. Look-alike web addresses and prices that don't add up, not only urgent wording.
- It listens to calls. A live caption on the phone, and a warning the moment a caller asks for a code or a "safe account".
- It never says "safe". When unsure: "not sure, don't pay yet".
flowchart LR
G[Meta glasses<br/>camera + mic] -->|compressed video<br/>or one photo| P[iPhone app<br/>Vision OCR, redaction]
P -->|one frame, only when<br/>there is something to read| S[Server<br/>rules + Gemini]
P -.->|offline| R[Rules on the phone]
S --> V[Verdict, reason, signs]
R --> V
V --> O[Spoken on the glasses<br/>card on the phone]
| Part | Stack | Read more |
|---|---|---|
| iOS app | SwiftUI, Meta Wearables Device Access Toolkit 1.0.1, Apple Vision, HEVC decoding, Kokoro-82M voice on device (FluidAudio), Apple speech recognition | ios/ |
| Server | Python, FastAPI, Gemini, English and Finnish rules, look-alike domain and price checks, redaction | api/ |
| Website | HTML, GSAP, a Blender-rendered 3D intro, an in-browser simulator with real server results | web/ |
- Permissions are asked for in plain words; refused, the app says so and the phone camera still works.
- No connection: the phone reads the text itself and applies basic rules.
- Glasses hot, low or disconnected: the app tells you, and you check with the phone camera.
- Speed and battery: the glasses send 7 frames a second; a frame leaves the phone only when there is something to check.
- Privacy: nothing is stored. Card, IBAN, ID, phone and email numbers are removed before text leaves the phone. Audio is never uploaded. Meta analytics are off.
- Ran on real Meta glasses on 14 Nov 2026: glasses video in, verdict and spoken warning out. Call check ran through the glasses microphone.
- ~2 s from frame to verdict (timed at 1.7, 2.0 and 1.7 s).
- 44 Finnish and English examples pass as a regression set. This is not an accuracy claim.
- Try it without glasses at secondlook.pgf.lv: the verdicts in the simulator are real server output on real photos.
ios/ iPhone app for Meta glasses (Swift)
api/ FastAPI server: rules, Gemini, eval set
web/ Landing page and Try it simulator
deploy/ nginx + systemd config for the server
presentation/ Submission text, sources, report, video
docs/ Images for this README
Quick start: the app in ios/README.md, the server in api/README.md. Secrets live in api/.env and ios/Config.xcconfig, which are git-ignored; copy the .example files to start.
- Show the alert on the Ray-Ban Display (built, not showing yet).
- Test with people of different ages.
- Pilot with a bank's fraud team or a consumer organisation.
![]()
Built at Junction 2026 for the Meta challenge "AI Glasses for Everyone, Everyday".
Made by Progresful (Emils Plavenieks, Jurgis Kacens, Krisjanis Mezulis) 2026

