← Back
Design SystemDesign OpsAIAmundi SuiteAsset Management

Unify — Amundi

Exalt × Amundi — AI-assisted design system management

Role Product Design ConsultantYear 2026–presentClient Amundi (via Exalt)
AI toolkitskillsagentsPredictthe impactAcceptthe kit updateCheckby propertiesFixon a branch
Unify →Public websitesInternal toolsWhite label
AI-assisted kit updates — every change is predicted, checked and fixed before it reaches the three contexts.

Overview

On assignment at Amundi, I contribute to Unify — the design system shared by public websites, internal tools such as Amundi Suite, and white-label products for institutional clients. I built a way to manage the design system with AI: keeping kits consistent, checking what ships, and moving product journeys forward on solid ground.

Challenge

One design intent has to hold across three contexts, each with its own users, brand themes and constraints. Every kit update ripples through every project file — and in regulated asset management, "it looks fine" isn't a check. The hard part isn't drawing components; it's guaranteeing nothing breaks as the system evolves.

Process

  • ·Built an AI toolkit for the design system with Claude Code and the Figma MCP: skills that know Unify, its usage rules and the product doctrine, and read-only agents that find a component, audit a screen or prepare a kit update
  • ·Industrialized kit updates across project files: predict the impact before accepting, check by properties afterwards, fix on a branch — a single pass covered around a hundred components
  • ·Audit by properties, never by eye — and formalized a verification method shared with the team
  • ·Contributed to aligning brand themes for white-label products, and to the design system documentation in zeroheight
  • ·Moved Amundi Suite journeys forward (global search, knowledge base) with one simple rule: everything is built from design system components, nothing is redrawn by hand

Results

  • ✓Kit updates became a repeatable, verifiable procedure instead of a risky manual pass
  • ✓A method that separates what the AI observes from what it concludes: reliable at describing a file's structure, but any "this component doesn't exist" claim must be proven against the source
  • ✓Journeys built on published components — a clean handoff to developers

Other projects

Wero
BNP Paribas — Payment integration
Mes Comptes V2
BNP Paribas — Full app redesign
Banque de France
Octave & Octave — Website redesign
Life Insurance
BNP Paribas — Web banking
Appart Hunt
Personal project — Apartment search tool