Description
- THE SITUATION: We sell premium beef jerky in Paraguay via WhatsApp - cash on delivery, cold traffic from Meta ads, Spanish ("vos"). Our WhatsApp sales bot ALREADY EXISTS and is talking to real customers right now. You are not building from scratch. You are making an existing, working system sell better. You will get a WhatsApp number to chat with our current bot yourself before you commit to anything, so you know exactly what you're taking on. THE STACK YOU'LL WORK IN: (no platform migration - you adapt OUR code) - Custom Python backend (FastAPI), deployed via GitHub Actions - Deterministic state machine for the sales flow, prices and orders (the AI never invents numbers — that stays) - LLM layer (Anthropic Claude API) for understanding and routing - PostgreSQL, WhatsApp gateway, decision logging, full automated eval suite (every behavior change must keep ~230 test cases green) - THE PROBLEM: Reply quality in the messy real world: customers go off-script, stall ("thanks, I'll let you know"), ask sideways questions - and the bot doesn‘t always pull them back into the scriptforward to the close. Your job: make the bot sound human, never break character, recover off-script customers, and advance every conversation toward the sale - inside our existing codebase. - THE #1 REQUIREMENT WE WILL TEST YOUR BOT You must have already built a WhatsApp sales bot for another company that is live and closing sales. In your proposal, include a WhatsApp number where we can chat with a bot YO