Grey-Workers-Togo/dodzi

Ai Spam & Fraud detection # Dodzi **Prends le temps.** — Anti-fraud message analysis for messaging apps first, email later. **Status:** pre-MVP, specification complete, ready to build **Owner:** Grey **Last updated:** 2026-08-15 > Building here? Read `CLAUDE.md` first — it has the non-negotiable > constraints, the stack, the repo layout and the build order. --- ## What this is A consumer anti-fraud assistant that reads the messages a person actually receives — WhatsApp, SMS, Telegram, Messenger, later email — and tells them, in their own language, whether a message is trying to defraud them and **why**. The wedge is deliberate: **most scams in our target market contain no URL and no attachment.** They are pure social engineering in French, Ewe/Mina and pidgin, usually ending in a mobile-money transfer or an OTP handover. URL-reputation tools — VirusTotal and everything built on it — are structurally blind to those. *Dodzi* is an Ewe name meaning patience. Every scam works by removing it. See doc 13 — the name still needs native-speaker and trademark confirmation. ## Two things that changed on 2026-08-14 1. **Meta announced WhatsApp Scam Alert** — on-device ML over linguistic and conversational signals, non-contacts only, free, native. A substantial part of what this product was going to be. Docs 01 and 09 are revised accordingly. The surviving position is **cross-channel + compromised-contact scams + local scripts**. 2. **The build order changed.** The WhatsApp forward-to-check number ships *before* the Android app (ADR-0014) — real users and a growing corpus in week 3, with nothing blocked on Play review or the IPDCP filing. ## Documentation map | Doc | What it answers | |---|---| | 01 — Product brief | Who it's for, what it does, why it still wins | | 02 — Architecture | How the system is put together | | 03 — Detection engine | How a verdict is actually computed | | 04 — Data model | What we store, and what we deliberately don't | | 05 — API spec | Contract between clients and …