Pump · 2026
Atelier: when a Design System stops being a library and becomes infrastructure
When AI started building interfaces, we realized the Design System had to stop being a library and become infrastructure.
- My role
- Staff Product Designer
Design Engineer - Company
- Pump
- Year
- 2026
Summary
As the platform grew, every new feature widened the gap between design and implementation. Duplicated components, inconsistencies, and outdated documentation were symptoms of a bigger problem: the system was not keeping up with the pace of the product.\n\nAtelier was born to fix that. Instead of maintaining just a component library, we built an infrastructure where design, documentation, and code evolve together.
Challenge
The challenge was never to build more components. It was to build a system able to scale on three fronts at once:
- development speed;
- team autonomy;
- consistency across experiences.
With the arrival of AI a new consumer of the platform appeared: agents also had to understand and build interfaces predictably.
Instead of solving each problem separately, we chose to build one architecture able to carry them together.
Traditional DS vs Atelier
Changing the role of the Design System
Organize components
Make product evolution predictable
Library separate from the product
Inside the product itself
Design and code need to be reconciled
Code and documentation are born together
Depends on human review
Automatically guaranteed by the system
Designers and devs build screens
Teams, clients, and agents build on the same system
Each brand requires adaptations
White-label configured via tokens
Reusable library
Platform capable of evolving
Approach
- 01
Production from day one
We built every component to run in production from day one. The goal stopped being to represent interfaces and became to generate reliable ones.
- 02
Rules before the interface
Components came out of tokens, contracts, and shared rules. The interface became a consequence of those decisions, not their origin.
- 03
Every delivery strengthens the system
Every new use case reinforces the system instead of creating exceptions. That way consistency grows along with the product.
- 04
White-label as configuration
Brands, themes, and behaviors are configured through tokens, allowing global evolution without touching components.
Architectural decisions
Layered tokens
We organized tokens in layers to separate global decisions from the ones specific to each brand. That made customization predictable without adding complexity.
The system audits itself
The system itself catches inconsistencies before they reach the product. Reviews no longer depend on manual inspection alone.
Atelier + Workbench
Components, documentation, examples, and tests all live in the same environment, keeping design and implementation in sync.
Extensible by customers
Customers can build their own components and patterns on the same infrastructure, without losing compatibility with the core system.
Agents as first-class consumers
AI stopped being only a tool for creating and became a consumer of the system itself. Components, contracts, and documentation were structured so agents can understand and generate interfaces with the same predictability expected of people.
Atelier, live
Part of Atelier can be tried out here. The color foundation shows how two brand decisions generate the product's entire palette. The Button page shows how those rules reach a component, from anatomy to states and accessibility.
Foundation
Component
Results
109 components
Shared foundation for the whole platform
1,165 consumption links
Real connections between Atelier and the product
65.8% reuse
Specs used by two or more consumers
21% of stories
Deliveries that also strengthened the system
Atelier's biggest result wasn't reaching 109 components, but building infrastructure able to reveal where the system generates the most leverage, where decisions still get repeated, and which investments should come next.
Atelier started evolving alongside the product: its components are consumed by the features, new stories strengthen the system, and real usage data guides its evolution.
It doesn't just organize what's already been built. It also helps decide what should be built next.
I built an infrastructure, measured its use, and used real data to guide its growth.
Next project
Building a car-sharing platform for São Paulo