- Role
- Design, engineering, deployment
- Year
- 2026
- Status
- Live · public demo
Overview
Most small Shopify brands sell on WhatsApp long before they have a support team. Messages arrive at midnight, on weekends, in three languages — and every unanswered one is a customer who buys somewhere else.
ReplyFlows puts an AI agent on that channel. It connects to the store's catalog, answers with real product data, looks up orders, and hands the conversation to a human when it should. Anyone can message the live demo agent without signing up.
The problem
Shoppers ask the same questions over and over — is it in stock, what size, how much is shipping, where is my order — and each one needs a human to open the admin panel and check.
Generic chatbots make this worse: they answer confidently from nothing, invent prices and quote products the store does not carry.
What I built
The agent is grounded in the merchant's own data. Products, prices, inventory and order status are pulled from the Shopify Admin API and kept in sync, so answers reflect what the store actually sells right now.
Instead of a scripted flow tree, the agent resolves intent and calls the tool it needs — catalog search, product lookup, order status — then replies on WhatsApp in the customer's language. When it lacks a grounded answer it asks a clarifying question or escalates, rather than guessing.
How it works
The path a single request takes through the system.
- 01
Connect the store
Shopify authorization syncs products, variants, prices, inventory and orders into the agent's own store, with webhooks keeping it current.
- 02
Index the catalog
Product data and store policies are normalized and indexed so the agent can retrieve the right item for a vague, human question.
- 03
Receive the message
Inbound WhatsApp messages hit a webhook, are matched to a conversation, and carry the thread's history into the reasoning step.
- 04
Reason and call tools
The model decides between catalog search, product detail, order lookup or policy answer — every response is built from retrieved data, not from memory.
- 05
Reply and record
The answer goes back on WhatsApp with product cards where useful, and the conversation is stored for the merchant's dashboard.
Capabilities
Catalog-grounded answers
Availability, price and specification questions answered from live store data.
Private order lookup
Order status resolved per customer, without exposing anyone else's data.
Multi-language replies
The agent answers in the language the customer writes in.
Merchant dashboard
Conversations and activity collected in one view instead of a phone inbox.
Human handoff
Conversations escalate to the team when the agent is not the right answer.
Public live demo
A demo store and number let anyone test the agent before connecting anything.
The product
Screens captured from the live site.
Technology
- TypeScript
- React
- Node.js
- PostgreSQL
- Prisma
- WhatsApp Business API
- Shopify Admin API
- LLM tool calling
- Docker
- VPS deployment
What changed
- A working commercial product — public site, live agent, onboarding and pricing — rather than a prototype.
- Merchants can test the agent by sending a real WhatsApp message, with no signup and no store connection required.
- Answers stay tied to catalog and order data, which keeps the failure mode “I need to check” instead of a confident wrong price.
Next case study
Nexora