A small demo — application to Dotworld

A back-office to generate and run AI micro-SaaS products in minutes

From the back-office, a new product is created by filling in a form, and each one is run on data — funnel, AI cost, margin — up to the decision: scale it, or cut it.

Public demo: payment and email are simulated here.
Built with Claude Code, specs and tests included: Discover the making-of

Next.js app
1
dynamic sub-apps
3
specialized agents
8
merged pull requests
98
passing tests
1,833

The back-office, in pictures

/admin
Back-office portfolio: four studio-wide KPIs over 30 days and the product table with its status.
The portfolioEvery product at a glance: visits, conversion, revenue, AI cost, margin, and the "scale it" or "cut it" suggestion.
/admin/products/lettre-pro
LettrePro product sheet: revenue, ARPU, AI cost, margin per generation and conversion funnel.
The product sheetEach product's funnel, from visit to purchase, and its visits and purchases curve.
/admin/products/new
Product creation form, Landing & SEO step, with the landing preview on the right.
Creating a productA 7-step form, with a live preview of the landing page as you type.

Why this project

Several SaaS products running in parallel, a modern stack, agentic development at the core of the process, self-funded and profitable from year one, grown from 1 to over 50 people in four years: a model that has proven itself. This demo is my way of concretely showing the wish to learn this model and contribute to it, by exploring its core business at my own scale.

The studio's products

Each one is a configuration of the same foundation, served on its own URL: click a card to try it live.

How it's built

"We develop with AI, not alongside it": this repo shows the method as much as the product.

  1. 1

    Specs before code

    One spec per feature, written before the implementation: behaviour, edge cases, acceptance criteria.

  2. 2

    TDD across the whole implementation

    Every acceptance behaviour is tested before it's coded, red then green.

  3. 3

    One spec, one multi-angle agentic review, one PR

    Every feature goes through several specialized review agents — correctness, Next.js, database, security — before being merged.