Take cake shop orders on WhatsApp in Nepali and English with an n8n agent

Kushal Sapkota, an independent developer in Nepal, built an ordering agent for a cake shop client across five n8n workflows and 98 nodes. Messages come in from WhatsApp (through the 360dialog API) and a website chat, often mixing Nepali, Roman Nepali and English, as text, voice notes, photos or payment screenshots. Each one is deduplicated, transcribed or fetched, and checked against a pause flag for staff takeover. An AI Agent node with memory and five SQL tools (cakes, occasions, delivery zones, time slots, payment methods) returns a reply plus structured fields. Code then rechecks the owner's feature switches, validates items and dates and computes totals and fees in SQL. A vision model reads payment proofs. The bot replies with photos and a payment QR, saves the order in Postgres and alerts the owner on Telegram.

The author says it has been running in production for the client. Things that broke: the website chat went quiet for about an hour after a regenerated workflow was left inactive, a test left the bot stuck in English so a Roman Nepali customer got an English reply, and Meta rejected Nepali-script documents twice during WhatsApp business verification. No order volumes or time savings are given.

The model never decides money. It cannot mark an order paid; a payment screenshot only flags the order for the owner to confirm, and prices come from the database, so an admin-panel change applies on the next message. Tests run on a copy of the workflow with send nodes off and fake numbers. Check every workflow's active flag after each deploy.