Designing Trust Into a Conversational AI
role
Founding Designer
Collaborated with
Engineers, CPO, CEO
Timeline
Mar 2025 - Present
Project Summary
A privacy-first conversational AI assistant for personal finance. I was brought in to build features. I ended up questioning whether the product was working at all — and that turned out to be the most valuable thing I did.
Fig: Final design I shipped
OVERVIEW
A vault for your financial life
An early-stage startup building a conversational AI assistant for personal finance. The architecture was genuinely novel - user data lives in a personal vault they own completely. Banks, documents, manually entered fields - all of it. The AI reasons from it. No third party touches it.
The privacy promise was real. Whether users would ever feel it? - that was the open question.
Fig: The grim reality of manually searching 160+ documents and storing them locally.
BUILT TO SHIP
What I Was Brought In To Do
Build features. The team was small, moving fast, and needed design and engineering bandwidth.
My first problem: the product pulls data from multiple sources — bank accounts, documents, connected apps, manual entries. As users connect more, the knowledge graph grows. How do you display that in a way that's understandable, manageable, and scalable?
That became the profile UI. Tabbed by how users think about their own life, not how data is structured in the backend. Every field shows where it came from. Auto-save. Collapsible cards for data that can have multiple instances — multiple addresses, multiple accounts.
The design problem was real and I solved it.
But I kept asking a different question on the side.
Fig: The highlight cards in the dashboard and the questions they answer for the user
SIDE HUSTLE
Nobody had watched a real user try this
Nobody had watched a real user try this product end-to-end. Not once.
I kept pushing for user testing while features were being built. It wasn't the priority. So I did my own thing — I grabbed a friend, sat him down, gave him a real task.
How much should my holiday gift budget be?
He got through onboarding. Reviewed what data Charlie could see. Approved it. Waited.
Got a generic answer. Switched to ChatGPT.
Not because of a bug. Because the privacy ceremony cost more than it gave. By the time he finally got an answer, he'd already lost faith it would be worth it.
Fig: State 1: While on Home Page (i.e. the Lease Management page), Clara answers any question about all leases
Fig: State 2: While inside a specific lease, Clara answers specifically questions related to that lease
That one session changed how I thought about the entire product. The team had been designing around privacy as the headline - the thing that makes Charlie different. But users don't open the app for privacy. They open it for the answer. Privacy is what makes the answer trustworthy - it should be invisible infrastructure, not a toll booth.
I wrote it up and brought it to the product lead.
It reframed how the team talked about the product.
TESTING
Four User sessions. One thread.
Eventually I got the bandwidth to run proper sessions. Four of them across mobile and desktop. I wasn't testing UI. I was testing whether the core value proposition was landing at all.
Each session peeled back a different layer of the same problem:
Fig: Versions I thought of before going with the version 2. I chose version 2 because: The 'Snapshot' being clickable is more prominent
That last one was the sharpest. The bank connection flow (designed as the convenient option) read as the riskier one. He'd rather type his details in manually than have a tool pull them automatically.
That's not a UX fix. That's a product decision.
THE REFRAME
Privacy isn't the product. The answer is.
The throughline across all four sessions: users weren't uncomfortable with the privacy model. They were uncomfortable with how it showed up.
The team had been designing around privacy as the headline — the thing that makes this product different. But users don't open the app for privacy. They open it for the answer. Privacy is what makes the answer trustworthy — it should be invisible infrastructure, not a toll booth.
That reframe changed everything that came after it.
Voice input stopped being a feature request and became an infrastructure decision. If audio touches a third-party server unencrypted, the privacy promise breaks. I worked with engineering to route transcription through a secure enclave — audio encrypted before it ever leaves the device.
Then I designed around it: idle, recording, processing, error. Browser permission flows. Cross-browser differences.
The mic button isn't just a component. It's the product keeping its promise under pressure.
Tasks — recurring financial check-ins, each running in its own persistent thread. The AI carries context forward. You don't re-explain yourself every time.
The hard design question wasn't the UI.
It was: what does it mean for an AI to act on someone's behalf when they're not there?
I scoped the full interaction model — cadence logic, missed runs, triggered conversations, thread continuity — and made the interventions explicit rather than invisible.
IN THE WEEDS
What I shipped while the strategy work happened
While all of this was happening, I was also in the codebase.
Mobile onboarding had a CSS overflow bug clipping nav on mobile — fixed with a custom cubic-bezier animation and safe area insets for notched devices.
AI thinking states needed fallback phrases during processing — moved the mapping to the backend SSE stream so the frontend doesn't break when the model changes.
Full WCAG AA audit on the profile page — contrast failures fixed with semantic color tokens.
None of it glamorous. All of it making the product feel more trustworthy.
Fig: Building this entire flow for GrowthFactor, literally taking it from 0 to 1, is the highlight of my life.
REFLECTION
The uncomfortable question is always worth asking sooner
I was hired to build features and I built them. But the most valuable thing I did in the first two weeks wasn't a Figma file or a component — it was sitting a friend down and watching him use the product.
That one session was worth more than anything else I shipped.
If I did this again, formal testing would happen in week one. Not week six.
The features weren't wrong. But we built them before knowing if the foundation held.
Validate first. Build second.
See how I designed the Deal Management & Lease Engines That Power Retail Growth