Kyrie-ML/kenya-scam-shield

# Kenya Scam Shield Kenya Scam Shield is a privacy-first, explainable risk scanner for suspicious SMS and WhatsApp messages in Kenya. It combines a lightweight text model with contextual scam rules, known-campaign matching, passive URL analysis, organization verification, and sender context. It is a decision-support tool, not proof of fraud. A low score is not a guarantee that a message is safe. ## What the user receives - A clearly labelled 0–100 automated risk estimate - A `low`, `suspicious`, `high`, or `very_high` risk level - A likely threat category, with uncertainty preserved when evidence is weak - Separate claim verification: `verified`, `unverified`, or `known_scam` - The strongest evidence, suspicious-link findings, and social-engineering techniques - Practical next steps and recovery guidance when the user may already have interacted - Explicit notices when ML or external reputation data is unavailable The interface is mobile-first, keyboard accessible, works without JavaScript frameworks, and includes safe example messages. ## Kenya-focused detection Preprocessing preserves phone numbers, amounts, URLs, and local terms while normalizing Unicode, zero-width characters, repeated punctuation, and common evasion such as `M-P3SA`. Rules cover English, Swahili, and code-switched patterns including: - M-PESA, Fuliza, Paybill, Till and Airtel Money impersonation - Fake loans and advance fees - OTP, PIN, password and account-verification theft - Fake jobs, grants, bursaries and student schemes - Prize and promotion fees - Guaranteed-return investment groups - KRA, HELB, bank and foundation impersonation Single words such as “M-PESA”, “loan”, “free”, or “scholarship” do not establish fraud. Contextual composites require combinations such as a loan promise plus an advance fee, or a prize plus a fee to claim it. Legitimate transaction notices, OTP warnings, personal money messages, and official opportunities are represented in tests and sample data. ## Ar …