Hackathon project · The app is in Rioplatense Spanish, built for people in Argentina.
The problem
Scams over WhatsApp and SMS are everywhere in Argentina: the fake tax agency (ARCA/AFIP) threatening to freeze your accounts, the "Mercado Pago" message saying your account was blocked, the parcel held by Correo Argentino, the classic "hi mom, I changed my number", fake Marketplace sellers, "bank support" asking for your PIN and crypto schemes that promise to double your money.
The messages keep getting more convincing — many are now written with AI — and the people most exposed are the ones with the fewest tools to spot them: older adults.
The solution
AntiEstafa AR protects people before, during and after a scam.
| Moment | What it does |
|---|
| Before paying | Checks the alias / CBU / CVU you were asked to transfer to and detects fake transfer receipts. |
| While analyzing | Combines AI with real technical checks on every link and highlights the fake part. |
| Afterwards | Drafts the police report, alerts your family on WhatsApp and reads the result out loud. |
Features
Analysis
- Text, screenshot, voice note, link or QR code. Screenshots are read with vision AI; voice notes are transcribed (with a local Whisper model, so it works even without a paid AI key).
- AI tuned for Argentine scams — detects manipulation signals such as false urgency, impersonation of agencies and banks, and requests for money or credentials.
- Real link checks — domain age (RDAP), SSL certificate, redirect chain and brand impersonation (Levenshtein distance against known brands).
- 0–100 risk score with a traffic light — 🟢 safe · 🟡 suspicious · 🔴 scam — and the deceptive part of the domain highlighted (
mercadopag0 vs mercadopago).
- "This is what the real one looks like" — shows how the impersonated company actually communicates.
- Built-in fallback engine — if there is no AI key or the free quota runs out, a rule-based engine for Argentine scams plus link checks keeps it working.
Before transferring money
- Alias / CBU / CVU check — validates format and check digits, and flags suspicious requests. It never claims an account is safe; it always recommends verifying through another channel.
- Fake receipt detector — upload the "transfer receipt" someone sent you and see what doesn't add up.
If you already fell for it
- Step-by-step recovery guide depending on what happened (sent money, shared a PIN, card or ID details, installed an app, extortion), most urgent steps first.
- Police report generator — a formal text with the message, links, signals, verdict and date; copy, download or print to PDF.
- One-tap 911 call with the report ready and a script of what to say.
- Share an alert with your family directly on WhatsApp.
Learn and stay up to date
- Current scams circulating in Argentina, with sources.
- Practice quiz: scam or real?
- Pandi, a friendly panda assistant that answers questions in plain language — never asks for personal data and never makes up phone numbers.
Accessibility
- "Grandma mode" — the result in two lines, large text, one action.
- Read aloud in Argentine Spanish voice.
- Installable app (PWA) with install instructions for every browser, an in-app user manual and a mobile-first design. No sign-up.
For businesses (offered through tecnoros.ar)
- White label — each client with its own name, logo and contact, configured via environment variables.
- API for WhatsApp bots (
/api/v1/analizar) with per-client keys, monthly AI quotas and replies formatted for WhatsApp.
- Usage metering per client and per month for billing.
- Embeddable button for any website with one line (
widget.js).
- Integration guide:
docs/INTEGRACION_TECNOROS.md.
Privacy & security
- Messages are not stored. "My analyses" history (last 10) lives only on the user's device and can be turned off and cleared.
- SSRF protection on link checks (private IPs and localhost blocked, timeouts).
- Per-IP rate limiting, 4 MB image limit and retries when the AI provider is overloaded.
Performance
- Link and AI checks run in parallel, max 3 links per message, 1-hour cache per domain and 4 s timeouts.
- Visible step-by-step progress. Measured: link checks went from ~0.8 s to ~0.3 s with caching, and slow redirects no longer block the analysis (previously up to 5 s).
How it works
text / image / audio ──► /api/analyze ──┬─► links ─► RDAP · SSL · redirects · brands (parallel, cached)
└─► Gemini (JSON with responseSchema)
│
combined score + verdict ─► UI
link ──► /api/link ──► RDAP · SSL · redirects · brands
alias/CBU + message ──► /api/transferencia ──► format validation + Gemini
receipt screenshot ──► /api/comprobante ──► Gemini vision
question to Pandi ──► /api/asistente ──► Gemini (short answers, fixed rules)
| File | Responsibility |
|---|
src/lib/links.ts | Link extraction and checks, brand Levenshtein, anti-SSRF, timeouts |
src/lib/gemini.ts | Prompt with Argentine scam context, validated JSON output, model fallback and retries |
src/lib/score.ts | Combines technical signals with the AI verdict |
src/lib/transferencia.ts | Alias, CBU and CVU validation (check digits) |
src/lib/comprobante.ts | Transfer receipt analysis |
src/lib/denuncia.ts | Police report text |
src/lib/pwa.ts | Manifest and browser detection for installation |
src/components/ | UI: analyzer, traffic light, grandma mode and each feature |
Stack: Next.js (App Router) · TypeScript · Tailwind CSS · @google/genai (Gemini with vision and audio) · Whisper tiny + sherpa-onnx (offline speech recognition) · RDAP · Vercel.
Quality: 100+ automated tests (Gemini and network mocked, so they don't consume quota), lint and build on every change.
Run locally
npm install
cp .env.example .env.local # add GEMINI_API_KEY (without it, the built-in engine is used)
npm run dev # http://localhost:3000
npm test
On Windows (PowerShell) use Copy-Item .env.example .env.local instead of cp.
1. Fake ARCA tax debt → 🔴 scam
ARCA INFORMA: Registrás una deuda de $187.430 en Ganancias. Para evitar el embargo de tus cuentas regularizá HOY en https://arca-regularizacion.com/pago antes de las 24 hs.
2. Fake Mercado Pago → 🔴 scam — mercadopag0 (with a zero) is highlighted
Mercado Pago: tu cuenta fue BLOQUEADA por actividad inusual. Validá tu identidad en http://mercadopag0-seguridad.xyz/validar o perdés tu saldo.
3. "Hi mom, I changed my number" → 🔴 scam — detected through social engineering, no links
Hola ma, cambié de número, agendá este. Estoy en un apuro, ¿me podés transferir $85.000 al alias pago.rapido.mp? Después te explico, no me llames que no puedo hablar.
4. Fake Correo Argentino → 🔴 scam — correoargentlno uses an "l" instead of an "i"
Correo Argentino: tu paquete está retenido por falta de pago de $1.250 de aduana. Abonalo en https://correoargentlno-envios.com/pago para recibirlo mañana.
5. Legitimate message → 🟢 safe
Hola! Te confirmo el turno con la Dra. Gómez el martes 28 a las 10:30 en el consultorio de siempre. Si no podés venir, avisanos por este medio. Gracias!
The demo video is recorded automatically with Playwright, walking through every feature:
npx playwright install chromium # first time only
npm run dev # in another terminal
npm run demo:video # creates public/demo/demo.mp4, portada.jpg and demo.vtt
It uses fixed sample responses so the video is always the same. Options: --url=https://antiestafa-ar-2.vercel.app to record the deployed version and --real to use the real AI. Requires ffmpeg.
Documentation
Documentation is written in Spanish.
Team
| Member | GitHub | Contribution |
|---|
| Manuel Angiulli | @Ridioortv | Core app: AI and vision analysis, link checks, scoring, rate limiting and Vercel deployment; tabbed redesign, link check, recovery guide and analysis history. |
| María Sofía Genta | @sofiigentaa | Alias/CBU check, fake receipt detector, police report, WhatsApp sharing, voice, examples, highlighted domain, "the real one", follow-up questions, history, installable app and performance work. |
Built during a hackathon with the help of AI coding agents (Claude Code) working in parallel on the same repository.
