One AI coach, three places to meet her.
Aimee is Financial Finesse's AI financial coach. I designed her front end and coded the prototypes and component library engineering built from. Then I merged the token layer and four components into the production SDK myself.
People asking AI about retirement need to know what they're talking to
And they need a human within reach. Aimee had to deliver both in three places, each with its own frame and front door.
The problem
Three homes: the logged-in Hub, client benefits sites, and Microsoft Teams. Three front ends would split the brand and triple the upkeep.
The bet
One system: a design-system spec, a locked library of blocks Aimee speaks in, and coded prototypes engineering builds from.
My role
Visual language, interaction model, the spec, coded prototypes, and production components. Platform engineers built the SDK.
Team
Platform engineers (the SDK, React and A2UI) · the coaching back end and model team
Built with
HTML, CSS, and JavaScript prototypes with live streaming · a DESIGN-SYSTEM.md spec · tokens and components in the React SDK
Status
Components merged into the production SDK · usage not published
Every answer is built from the same pieces
Every answer is built from typed blocks. A colored rail tells you what kind of card you're reading before you read it.
Honest about the AI, with a human one tap away
She's an "AI coach," never a "virtual assistant." When the AI isn't enough, a named, certified coach joins the same thread.
A standing disclaimer
"Aimee can make mistakes. Please double-check important answers." sits under every composer, where people read answers.
A correction on every reply
Thumbs up, thumbs down, and copy under each answer.
Words first, then cards
Cards and charts wait until the sentence finishes, so an answer never looks half-broken.
Restraint as a rule
Three blocks per turn at most, one primary action per card.
A design rationale, not a measured trust outcome. Stream retry, missing context, and correcting a user's wrong assumption live in the SDK and back end engineering owns; I didn't design those states.
The same coach, wherever she lives
The vocabulary never changes. The host decides the front door, the frame, and how much room the coach gets.
The user's words never shrink
In one build, the user's message drifted to 14.5px while Aimee's reply stayed at 16. Review against the prototype caught it, and I made the rule part of the token layer so it can't drift again.
I designed it in code, then shipped it myself
Prototypes, not mockups
Real HTML, CSS, and JavaScript with live streaming and state, so flows could be felt before production code.
A spec people use
DESIGN-SYSTEM.md set the tokens, the type, every component, and the accessibility hard rules.
A contract to build from
Engineering built the SDK from the component library, and we reviewed it against the prototype.
Then I shipped it
I merged the token layer and the empty state, conversations, icon, and card-shell components into the SDK between July 21 and August 17, 2026.
What shipped, and what isn't measured here
Surfaces designed from one component library (scope, not usage)
Documented design system and component library, public on this site
Components I merged into the production SDK, plus the token layer, Jul 21 – Aug 17, 2026
No active-user or trust-outcome figure for Aimee is published here. The platform's customers employ 1M+ people who are eligible for it: that's reach, not usage.
One coaching vocabulary for three hosts, with the AI named honestly and a human one tap away. Designed by one person, built with engineering, and finished in production code by the same designer.