The customer writes at 3am and nobody answers

An AI salesperson answers in their language, quotes real prices, takes the order.

A run through the cockpit, demo data: a customer question typed in the test chat, the answer with the real price, stock and product photo, the confirmed order, then the orders and the inbox.
Customer inbox, demo data: the escalation the agent handed over, then WhatsApp and Instagram threads, each tagged with its language.
Test chat, demo data: the customer asks the price in Darija, the agent answers with the catalogue price, the stock, the delivery and the product photo.
Orders and escalations, demo data: each order created by the agent with its status and total, and the conversation frozen over an unverifiable price.
The cockpit, demo data: the Shopify catalogue as source, the agent and its rules, the WhatsApp, Instagram and web widget channels.

The problem

Merchants lose orders when customers write after hours, and generic chatbots make it worse by quoting prices that do not exist.

What I built

A multi-tenant SaaS where an agent answers on WhatsApp, Instagram and a web widget, quotes real Shopify prices and stock, and creates orders. The code re-checks every price against the catalogue; an ambiguous match goes to a human.

Technical details

Hybrid search in SQL: pgvector HNSW and full text, fused by RRF
Merchant secrets in Supabase Vault, reachable by the server only
Meta webhooks verified on the raw body, replay-safe
Draft orders behind three anti-duplicate locks

Stack

TypeScript strict
Supabase Postgres + pgvector
Claude API
Meta Graph API
Shopify GraphQL
Stripe

The result

Checked by 3,398 automated tests, including ones that prove the margin stays at 55% or more at real AI model prices.

A similar project in mind?

Tell me about your business and when you need it. You get a proposal by email.

I want this system

More projects

ERP & CRM · Warehouse stationThe returned parcel goes back on a shelf, never into stockView the project