CyberSathi — Offline AI Security Analyzer
On-device scam detection for messages, emails, screenshots, QR codes and URLs — powered by a local Gemma model via Ollama.
CyberSathi analyses suspicious SMS messages, emails, screenshots, QR codes and URLs for phishing, malware and fraud — in English, Urdu or Roman Urdu. No cloud LLM is used for detection: a local Gemma model served by Ollama runs the analysis, and every scan stays on the machine after the initial setup. A deterministic rule engine handles the parts that must never be guessed — sender domain checks, linguistics, malware and OTP-request heuristics, and a blocklist of more than fifty known scam domains — while the model handles free-form reasoning. Scans persist to a local SQLite database through SQLAlchemy, and the React and Tailwind frontend includes a PWA manifest and service worker for offline use.
Key Features
- Message, email, screenshot, QR code and URL analysis endpoints
- Local Ollama runtime with the Gemma 2B model — no cloud LLM in the detection path
- Deterministic linguistic, sender, URL, malware and OTP checks alongside model reasoning
- Offline QR decoding in the browser with jsQR — no external service
- Local SQLite history through SQLAlchemy, with scan review and deletion
- Full English, Urdu and Roman Urdu language support
- PWA manifest and service worker for offline caching
Repository Topics
Challenges
Balancing a probabilistic model with deterministic rules so that a scam verdict is explainable, and keeping the whole pipeline usable on a machine with no internet connection.
Results
A complete offline-first security companion that pairs model judgement with hard rule checks and a blocklist — and says exactly which factors drove each verdict.