PideBot
A WhatsApp bot that takes the order and collects payment for the restaurant, 24/7.
Customers message the same WhatsApp number as always. The bot builds the order against the real menu, charges through Nequi, and drops it on the kitchen dashboard — with nobody sitting there watching the phone.
Buenas, me regalas 2 bandejas paisas 🙌
Listo. 2 × Bandeja paisa — $56.000. ¿Te lo confirmo?
Sí, confirma
The problem
Small restaurants take WhatsApp orders by hand: someone has to read, transcribe, add up the total, and chase the transfer. At lunch rush that falls apart. And pointing a loose language model at that conversation is a real risk — it can invent a dish or quote a price that doesn't exist.
Engineering decisions
The model never sets a price
The conversation engine uses tool calling to propose an order, but totals are always recalculated from the menu rows in Postgres. The order is only created on an explicit customer confirmation.
The webhook never does the slow work
Meta and Wompi webhooks enqueue onto Redis and answer immediately; a separate worker process makes the model call and the writes. Meta times out in seconds — the heavy work cannot live inside that request.
Multi-tenant from the schema up
Every business table carries tenantId. One API serves every restaurant, and isolation is a property of the data model rather than a convention someone has to remember on each query.
Automatic Nequi collection
Wompi integration: the charge is created, and the payment webhook is what advances order state. The restaurant stops chasing customers for a screenshot of the transfer.
Kitchen dashboard
A Next.js panel over the same tables the bot uses: notifications inbox, live orders, menu editing, payment status.
Stack
Next project
Tendero
Inventory, sales and barcode scanning for a corner store. All from a phone.