Replatforming Atelier Supply to headless Medusa

Atelier Supply is a direct-to-consumer apparel brand. They had outgrown a themed commerce platform: the storefront was slow on mobile, every design change fought the theme system, and a percentage of every sale went to the platform.

The problem

The store worked, but three things were capping growth:

• Mobile performance. Time to first byte sat above 900ms and the largest contentful paint followed it. On the traffic that mattered (paid social, mostly on phones) the bounce rate showed it.
• A design ceiling. The team wanted editorial product stories, custom collection layouts, and a configurator for made-to-order pieces. All of it meant fighting the theme.
• The fee. Revenue-share on top of the subscription had grown into a real number as the brand scaled: money going out for no additional value.

What we built

A headless architecture: a custom Next.js storefront on an open-source Medusa backend.

• Storefront: server-rendered product and collection pages, a custom collection layout system for the editorial content, and the made-to-order configurator as its own flow.
• Backend: Medusa with a custom module for the made-to-order pricing and lead-time logic, Stripe for payments, and Algolia for search.
• Migration: catalogue, customers, and order history mapped to the Medusa data model and migrated with a tested script. Every old URL either kept its path or got a redirect.
• Infrastructure: Medusa, Postgres, and the storefront on hosting in Atelier Supply's own accounts, with a CI pipeline and staging environment.

How it went

The build ran against a staging URL from the first week. The team reviewed the real storefront as it came together, and the migration script was run against staging repeatedly until the data was clean.

Cutover was done out of trading hours with a tested rollback path. It was not needed.

The result

Time to first byte dropped 71%, and mobile conversion followed, up 2.3× over the three months after launch. The revenue-share fee is gone; the brand now pays for infrastructure only, a cost that does not scale with sales.

The team owns the repositories, the database, and the hosting. When they need a new collection layout now, it is a small pull request their own developer ships.