drizzy765

**Problem.** Small sellers in Nigeria run their businesses inside Instagram and WhatsApp DMs, and fake transfer screenshots are one of their biggest losses. A "buyer" sends a convincing bank or OPay receipt, the seller ships, and the money never arrives. I nearly lost goods this way myself. On top of that, sellers spend hours answering the same "how much?" questions, often in Pidgin or Yoruba, and through voice notes. **What Sell-Safe does.** Sell-Safe connects to a seller's real Instagram account and runs their DMs, powered by Meta Muse: - **It sells in the buyer's language.** Muse understands English, Pidgin and code-switching, then replies with real prices pulled from the seller's own catalog. Sellers can import that catalog from a Google Drive link. - **It understands voice notes.** Pidgin voice notes are transcribed with Intron Sahara, then Muse understands them and replies in kind. - **It sends payout details safely.** When a buyer asks for the account number, the reply uses exactly the payout details the seller saved, so the AI can't invent an account number or change a price. - **It protects every payment.** Muse reads the receipt (amount, recipient, reference), but a deterministic verification engine, not the language model, decides the verdict: CONFIRMED, NOT FOUND or SUSPICIOUS. A screenshot alone never releases goods. - **The seller stays in control.** A live dashboard mirrors the shop's real Instagram DMs, with every chat, payment verdict and setting in one place, plus an Auto / Approve / Smart reply switch. **Real data and measured results (SellSafe-Bench).** - **Payments:** 20/20 cases passed, with zero fake receipts confirmed. The suite covers edited amounts, reused references and a tampered screenshot. - **Prompt injection:** 6/6 attempts blocked. - **Live test on a real Instagram shop (@sellsafe.testshop.ng):** a Pidgin voice note was transcribed and answered, and a synthetic fake OPay receipt was held and not confirmed. - **Muse vision catalog extraction:** 1/4 exact so far (n=4). We report this as partial; it's still in progress. We report the weak numbers too, and the demo shows a real bug we found and fixed live. **Where Meta Muse fits.** Muse powers the conversations, understanding and replies, and reads receipts and product photos. The verdict that protects the seller's money is never the AI's guess. **Stack:** FastAPI (Python) on Render, Next.js on Vercel, the Meta Model API (Muse), the Instagram Graph API, Intron Sahara for speech, and Supabase. **Live:** [sell-safe-three.vercel.app](https://sell-safe-three.vercel.app/) · **Code:** [github.com/drizzy765/sell-safe](https://github.com/drizzy765/sell-safe)