Building a Multimodal AI Agent
— Tripling Conversion.
Led 0→1 design through engineering handoff
As the sole designer in T-Mobile's AI Innovation Lab, I led the end-to-end design of its first agentic AI assistant, from concept and interaction model to the shipped product handed off to engineering. It is rolling out across web, iOS, and Android. I owned the design system, the conversational UX, and the evaluation framework behind the AI's responses.
Switching carriers is overwhelming
65% of customers called changing to T-Mobile a hassle, and even with digital support, 80% of new customers still ended up in a store. People didn't need more information. They needed guidance.
There's too much to figure out on my own.
Research participant · switching study
I just want someone to tell me what's right for me.
What we set out to do
Goals for the redesignAn assistant that guides, not a search box that responds
Voice-led multi-modality
As the conversation moves, the interface follows. Visuals, options, and structure surface only when they help. People talk to it. They don't manage it.
It already knows what you're looking at
The assistant sees the same screen you do, so you never have to re-explain. Say "compare these two" or "is this one on sale for me?" and it picks up the context instantly — then continues as an agent, surfacing the right options and starting the upgrade without you spelling out the details.
Clarify by tapping, not typing
Too many follow-up questions means typing out an answer to each one — tedious fast. So the assistant clarifies in a quick quiz format instead: tap-to-answer choices, the way Claude poses clarifying questions to sharpen a prompt. Same understanding, a fraction of the effort.
What the MVP test changed
I shipped an MVP, watched how people actually used it, then redesigned the weak spots the data exposed. Each pair reads left to right: the problem the data surfaced, then the fix I shipped.
I owned how good the AI actually was
A great chat UI means nothing if the answer is wrong. I built the evaluation and red-teaming that made the responses trustworthy.
Measure it before we ship it
I built a red-team set of 300+ real customer questions and scored every answer against ground-truth plan data — catching wrong prices, missing discounts, and invented benefits before a customer ever saw them.
Rules that keep answers honest
Every failure became a principle the model and the UI both follow — so the assistant stays accurate, admits what it doesn't know, and stays easy to read under pressure.
Real cases the evaluation surfaced — and how the refined response fixed them.
One assistant, three platforms, one system
A calmer path to a confident decision
Won executive approval to ship. The work scaled org-wide
and led T-Mobile to stand up a dedicated AI design team.
Soojin bridged our AI and Product organizations and set the standard for how design leads at the intersection of the two.
