← All workCase study 2 of 4
AI agent · SaaSOwn product · private beta

vWorker

A multi-tenant SaaS that puts every customer channel of an online seller in one inbox, and adds an AI sales agent that replies in Bangla, English and Banglish: finding products, quoting prices the server computes, and taking orders and bookings.

Role
Solo: product, architecture, build
Status
Private beta with a pilot shop
Tests
~420 integration tests
Evals
36/36 on real models
The problem

Sellers answer every customer by hand, in five places.

Online sellers in Bangladesh sell through Messenger, Instagram, WhatsApp and their website. Questions, orders and courier bookings get lost between chats and spreadsheets, and nobody can answer at 2 a.m.

What it does

One inbox, one agent, real orders.

  1. 01One inbox

    Every channel in one place, for the whole team, with handoff from the agent to a person.

  2. 02An AI sales agent

    Answers in Bangla, English and Banglish, finds products and quotes prices the server computes.

  3. 03Orders and couriers

    Takes orders and bookings, with Pathao, Steadfast and RedX courier adapters.

  4. 04A platform

    A public REST API with signed webhooks for each seller's own systems.

Architecture

Built to keep every seller's data apart.

ChannelsMessenger, Instagram, WhatsApp and web chat through the Meta Graph API
QueueBullMQ on Redis: every incoming message is a job
Sales agentLLM tool use: product search over pgvector, server-side prices, order and booking tools
PostgreSQL + RLSRow-level security isolates every tenant in the database itself
DashboardNext.js 16 inbox, orders and settings
Public APIREST with signed webhooks
Messages arrive as queued jobs; the agent acts only through tools; the database enforces tenant isolation.
Key decisions

Where the engineering went.

  1. 01

    The model never sets a price

    Prices come from the server. The agent quotes what a tool returns, so it can't invent a discount.

  2. 02

    Tenant isolation in the database

    Postgres row-level security, not just application checks, keeps one seller's customers away from another's.

  3. 03

    Evals against real models

    36 real-model evals cover the conversations that matter before a change ships.

Where it stands

In private beta.

real-model evals passing
36/36
cost per AI reply
1–7 paisa
orders taken end-to-end on a real Facebook page
Live

Pre-launch: no revenue or customer counts are claimed.

Screenshots

Screenshots of the demo business only. A product demo is available on request.

TypeScriptFastifyNext.js 16PostgreSQL + RLSDrizzle ORMBullMQ / RedispgvectorLLM tool useMeta Graph API