A multi-agent pipeline that ships software from planning to production.

15 specialized AI agents coordinated by a Python orchestrator. A text brief goes in. A deployed, tested, production-ready application comes out.

6 sites deployed
3–5h per site
15 agents
5 quality layers
How it works

A Python orchestrator dispatches specialized agents through a deterministic sequence of phases. Each agent operates on the codebase via Claude Code, produces a pull request, and hands off to the next stage. Quality gates are deterministic wherever possible: the LLM writes code, but the checks that catch defects are structural, not probabilistic.

PHASE 01Planning
Reads the text brief. Generates architecture docs, API contracts, design prototypes, copy deck, and a full set of GitHub issues.
PM Architect Designer Copywriter
PHASE 02Infrastructure
Spins up local Supabase via Docker. Scaffolds database schema, migrations, and row-level security policies.
PHASE 03Development
Backend, frontend, content, SEO, and AI-agent tasks dispatched in parallel. Each task produces a pull request. Every PR is reviewed by a dedicated reviewer agent before merge.
Backend Frontend Content SEO AI-Agent Reviewer
PHASE 04Quality gates
Five verification layers: deterministic static analysis (TypeScript, imports, i18n parity, dead code), LLM-based code review, database audit, security audit, and Playwright E2E tests. When the fix loop stalls, a forensics agent diagnoses root cause.
QA-Static QA-LLM DB-Review Security E2E Forensics
PHASE 05Deploy
Supabase Cloud provisioning, Vercel production deploy, smoke tests on all routes and locales.

Each run produces: architecture documentation, API contracts, database schema with row-level security, React components, translation files, Playwright E2E tests, security audit, AI-generated illustrations, and a Vercel deployment.

Pipeline output

Every site below is live. Each was generated from a text brief describing a fictional business. Different industries, different page structures, different locales.

Durand & Associés
Durand & Associés
Law firm, Paris. Bilingual, 5 pages, 4 practice areas, attorney profiles, contact form.
bilingual multi-page form
Visit →
Atelier Verne
Atelier Verne
Architecture studio, Bordeaux. Bilingual, 6 project detail pages, bioclimatic design focus.
bilingual multi-page
Visit →
Le Comptoir d'Émile
Le Comptoir d'Émile
Restaurant, Marseille. Bilingual, seasonal menu with prices, chef profile, gallery.
bilingual
Visit →
La Mie Ancienne
La Mie Ancienne
Artisan bakery, Paris. Product catalog, opening hours, Too Good To Go link.
5h 3 passes
Visit →
Toitures Martin
Toitures Martin
Roofing company, Nantes. 5 services, past projects, quote request form with validation.
form
Visit →
Cabinet Morel
Cabinet Morel
Physiotherapy clinic, Lyon. 5 specialties, 3 practitioners, Doctolib booking.
3h12 1 pass
Visit →
IN PROGRESS Facture.simple Invoicing SaaS for freelancers. Bilingual, authenticated member area, persistent data, Supabase backend.
AI-assisted development

Kiroku

A separate project, built with a different workflow. Not the multi-agent pipeline above, but AI-assisted development throughout the entire process.

Kiroku is an open-source trading journal. The previous version had taken months to build. This version was rebuilt from scratch in 5 days: same scope, same features, same technical complexity (React, Supabase, authentication, persistent data).

The difference: AI as a development partner at every step, from architecture decisions to implementation to testing.

Previous version
months
hand-coded
Full rebuild
5 days
AI-assisted
GitHub →
Kiroku trading journal
About

15 years shipping software: developer, project manager, production director. I built this pipeline because I believe AI changes how software gets delivered. Not by writing code faster, but by orchestrating the entire delivery process: planning, development, quality, and operations.

LinkedIn →