AI powered scam call detector. Free forever. 100% on device. Zero data collection.
ScamReaper listens to the transcript of a call, runs it through a TensorFlow Lite model that lives entirely on your phone, and tells you whether the caller is a scam — in under a second, without sending a single byte to a server.
No accounts. No cloud. No ads. No tracking. No subscriptions. Ever.
⚠️ This project is under active development. It is NOT yet available on the App Store or Google Play Store.
What exists today:
- ✅ Complete UI built and running on simulator
- ✅ TFLite AI model trained locally
- ✅ iOS 26 Call Screening integration coded
- 🛠 Android live call monitoring in development
- ✅ Security system implemented
What still needs to happen:
- 🔑 Apple Developer Account for real device testing
- 📱 Real device testing on iPhone with iOS 26
- 🔗 Native CallKit extension proper linking
- 🤖 Android build and testing
- 🚀 App Store and Play Store submission
Platform plans:
- iOS: Pre-call screening using iOS 26
- Android: Full live call monitoring during calls (Android gives more system-level access)
If you want to help get this to a shippable v1.0, see Contributing. Real-device testers and Android engineers especially welcome.
Real screenshots coming soon. Placeholders below.
| Home | Call Detail | Stats | Security |
|---|---|---|---|
 |  |  |  |
📞 Incoming call
│
▼
📝 iOS Call Screening delivers a transcript
│
▼
🌐 Language detector
├── Not English? ──────────▶ 🟡 YELLOW (always — we won't guess)
└── English ──┐
▼
🛡 Anti-bypass check
├── Keyword stuffing / robotic / repeated script? ──▶ 🔴 RED (forced)
└── clean ──┐
▼
🤖 TensorFlow Lite bag-of-words classifier (on-device, ~5 ms)
│
▼
🎯 Confidence tier
├── < 0.70 ──────────────▶ 🟡 YELLOW (unsure)
├── 0.70 – 0.85 ────────▶ verdict + "Low Confidence" tag
└── ≥ 0.85 ──────────────▶ verdict
│
▼
🔔 Smart notification
├── 🔴 SCAM BLOCKED [View Details] [Block Number]
├── 🟢 Legitimate [Accept] [Decline] [Ask More]
├── 🟡 Unsure [View] [Call Back] [Dismiss]
└── 🌐 Unknown Language [Listen] [Block] [Dismiss]
│
▼
💾 Stored locally (last N calls). Nothing ever leaves the device.
- 🔴🟢🟡 Three-class on-device verdict — RED / GREEN / YELLOW
- 🤖 TensorFlow Lite model (~7 KB, <5 ms inference on an iPhone)
- 🔔 Smart notifications with action buttons (Accept / Decline / Ask More / Block / View)
- 🌐 Language awareness — Spanish, Hindi, Telugu, Mandarin, French, Arabic, Russian detected and flagged YELLOW so the English-trained model never judges them
- 🛡 Anti-bypass detection — keyword stuffing, hidden payloads, robotic speech, and repeated scripts from the same number force RED
- 🔒 Model integrity verification — SHA-256 fingerprint of the
.tfliteis pinned on first run; app alerts if tampered - ⏱ Timing-attack prevention — every verdict takes the same wall-time (no latency side channel)
- 📊 Statistics dashboard — calls screened, scams blocked this week/month/today, time saved, 7-day trend
- 🛑 Security audit log — last 100 events (integrity checks, bypass attempts, fallbacks) stored locally and viewable in-app
- 😄 Rage Mode (opt-in) — on RED calls the AI suggests a harmless distracting reply in one of six languages to waste the scammer's time. Strictly non-malicious, off by default.
- 👋 Friendly onboarding — grandma-proof copy, no jargon
ScamReaper will never:
- Make a network request related to call analysis.
- Upload transcripts, phone numbers, or verdicts.
- Use advertising, tracking, or analytics SDKs.
- Require an account, login, or subscription.
Everything is computed with on-device code you can read in this repo.
| Layer | Tech |
|---|---|
| App | React Native + Expo SDK 54, expo-router |
| Inference | TensorFlow Lite Swift 2.14 (on-device) |
| Native bridge | Swift + Obj-C, RCT_EXTERN_MODULE |
| Integrity | CryptoKit SHA-256 fingerprint pinning |
| Storage | AsyncStorage (local, unencrypted — no secrets) |
| Notifications | expo-notifications with categorized actions |
| Training | Python 3.11 + TensorFlow 2.14 (ml/train_model.py) |
| License | MIT |
Requirements: macOS with Xcode 15+, Node 18+, CocoaPods, Python 3.11+
# 1. Clone
git clone https://github.com/<your-handle>/ScamReaper.git
cd ScamReaper
# 2. JS deps
npm install
# 3. iOS native depscd ios && pod install &&cd ..
# 4. Train the AI model (the .tflite is gitignored — you generate it)cd ml
python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python train_model.py # writes scam_detector.tflite + vocab.jsoncd ..
# 5. Copy model into the iOS bundle (first time only)
cp ml/scam_detector.tflite ios/ScamReaper/
cp ml/vocab.json ios/ScamReaper/
# 6. Run
npx expo run:iosThe classifier is a 3-class bag-of-words network. Dataset lives in
ml/training_data.py — add your own labeled examples and retrain:
cd ml
source venv/bin/activate
python train_model.pyOutputs:
scam_detector.tflite— the model iOS loadsvocab.json— the shared vocabulary (Python and Swift both tokenize with it)
Training tips
- Keep
REDexamples diverse — include IRS scams, fake bank fraud, tech support, lottery, romance, and "you owe us money" variants. - Balance the classes — roughly 40% RED, 40% GREEN, 20% YELLOW works well.
- Test on your own examples before committing. The toy dataset is intentionally tiny.
Pull requests are welcome. See CONTRIBUTING.md for details. A few principles:
- No network calls for anything related to screening. The privacy promise is load-bearing.
- No third-party analytics or ad SDKs. Ever.
- Multilingual respect. If you don't have training data for a language, return YELLOW — don't guess.
- Security over speed. A ~200 ms verdict is fine. A 20 ms verdict that leaks information is not.
Run npx tsc --noEmit before opening a PR.
Please do not open a public GitHub issue for security bugs. Use the Security Vulnerability template to open a GitHub Security Advisory instead, or email the address configured on the repo. See docs/SECURITY.md for the full process and SLA.
ScamReaper is for everyone who has ever been scared by a phone call — and especially for the people in our lives who are most targeted. If this project helps one grandmother, it was worth building.
We especially welcome:
- 🧠 AI/ML engineers (model quality, multilingual data)
- 🛡 Security researchers (anti-bypass, integrity, timing)
- 🌐 Multilingual contributors (scam patterns in your language)
- 📱 iOS / Android developers (CallKit, Call Screening, notifications)
Star ⭐ the repo, tell a friend, translate the UI, add a scam pattern you have seen. Everything contributed stays on-device, forever free.
MIT — see LICENSE. Copyright © 2026 ScamReaper Contributors.