01Own product
AI Sales Desk
An omnichannel inbox with an AI sales assistant for online sellers — architecture, design and engineering by DEVS KITCHEN.
- Client
- In-house product
- Services
- Product strategy, System architecture, UI/UX design, Web development, AI integration
- Platforms
- Web, Messenger, Instagram, WhatsApp
- Status
- In development
The challenge
Online sellers often take orders through a Facebook page, Instagram, WhatsApp and their own website at the same time. Customer questions arrive in four different apps, and the same few questions — price, size, stock, delivery — fill the day.
They need one place to answer everyone, and help with the repetitive questions — without an AI that invents prices or promises stock that is not there.
Our approach
We brought every channel into a single realtime inbox. Messages from the website widget, Messenger, Instagram and WhatsApp share one write path, so each is stored, de-duplicated and delivered the same way.
The AI assistant is deliberately constrained. It answers only from the shop’s own documents and catalog, uses a small set of strictly validated tools — search products, check stock, draft an order — and hands the conversation to a person when it is not confident. Prices are always calculated on the server, never by the model.
Each business’s data is isolated at the database level with row-level security, and every phase shipped against written acceptance tests — including concurrency tests for stock and duplicate-webhook tests for messaging.
What we built
- 01
Realtime omnichannel inbox
Assignment, tags, notes, priorities and live typing across website chat, Messenger, Instagram and WhatsApp.
- 02
Grounded AI assistant
Retrieval from the shop’s own knowledge base with pgvector, strict tool validation and hand-off to a person.
- 03
Commerce in the conversation
Products, variants, inventory and orders, with server-side pricing and duplicate-safe order creation.
- 04
Embeddable chat widget
A 57 KB widget isolated in shadow DOM, added to any website with one snippet.
- 05
Automation rules
Versioned, loop-protected rules that tag, assign, prioritise or reply when a message or order matches.
- 06
Multi-tenant by design
Row-level security on every table keeps each business’s data separate from every other.
Inside the product
Catalog — the prices and stock the team and the assistant quote 
Website chat widget — the customer’s side, on a phone 
Orders — created from the conversation, with full history
Stack
- Next.js
- React
- NestJS
- TypeScript
- PostgreSQL
- pgvector
- Prisma
- Redis
- BullMQ
- Socket.IO
Where it is now
The core platform is built and tested. Website chat works end to end, and the inbox, AI assistant, commerce, billing and automation pass their automated test suites.
Messenger, Instagram and WhatsApp are code-complete and waiting on Meta’s app review. The AI assistant will be evaluated with a production model before launch.