I made Etsy's AI feel human enough to trust.
Product Designer embedded in Etsy's AI recommendation team across buyer and seller experiences. I executed end-to-end UX research, redesigned transparent recommendation systems for buyers, and built AI-powered listing tools that gave sellers control rather than taking it away.
JUL – DEC 2025 · NO PROPRIETARY SCREENS SHOWN · NDA PROTECTED
01 · THE MOMENT IT STARTED
A plain document. An exhilarating problem.
THE DOCUMENT
The UX research brief was plain and boring. But what was written in it added so much depth to my day I felt exhilarated. Users were having problems with the new AI system. I was not surprised. I was excited.
THE TESTING ROOM
I became a silent tester. I watched Liz guide our buyer persona — sounding exactly like an AI. The user opened up with every question. Gave more details than necessary. Then he said: "I don't know why I'm seeing this." I mentally chuckled. I didn't know then this was just the tip of the mountain.
"I feel the system is deciding for me."
— Seller participant, research session. Three hours of active listening. I knew we had a big gap.
02 · THE REAL PROBLEM
It wasn't a bad AI. It was uncertainty spreading like a virus.
HOW I FOUND THE ROOT CAUSE
After synthesis across buyer and seller interview sessions, alignment with internal stakeholders, engineering teams, and the research team — the root cause became clear. It was not inconsistent design. It was not confusing navigation. It was purpose that was not transparent enough. In the new age of AI and personalisation, the very core of freedom was missing.
BUYERS
Wanted discovery — but not manipulation. The AI was surfacing products but not explaining why.
SELLERS
Wanted efficiency — but not loss of creativity. The AI was suggesting without giving control back.
AI SYSTEMS
Worked statistically — but not emotionally. The gap between algorithmic logic and human feeling was the real design problem.
03 · THE WRONG ASSUMPTION
I assumed explaining once would be enough.
I added microcopy. Simplified interactions. Cleaner flows. I tested it. It did not solve the problem.
WHAT I ASSUMED
"If we explain AI once, users will understand it."
WHAT I LEARNED
They needed ongoing transparency — not a tooltip. I circled back to the core pain points. I knew the only way forward was testing hypotheses built directly on what buyers and sellers had told me.
04 · SOLUTION
I was on a mission to visualise how AI would show up — not just work.
FOR BUYERS — TRANSPARENT RECOMMENDATION CARDS
Designed cards that answered the unspoken question: "Why am I seeing this?" Instead of hiding the system, I surfaced it — "Recommended based on your saved favorites" / "Trending among similar buyers." Tested prototypes again and again until hierarchy, consistency, and navigation felt right.
FOR SELLERS — AI THAT SUGGESTS, NEVER DECIDES
They didn't want magic. They wanted control. AI-powered bulk listing tools where AI could suggest — sellers always decided. Edit options on every suggestion. Educational microcopy explaining how Etsy's AI learns from tags, titles, categories. Every flow: editable, reversible, human.
SYSTEM LEVEL — DESIGN SYSTEM EXPANSION
Expanded Etsy's design system with modular, accessible AI components ensuring consistency across buyer and seller experiences. Documented every decision in Notion — not just what we chose, but why — so teams could move fast without re-litigating intent. Trust breaks when patterns don't repeat.
SEARCH CONSTRAINT — BALANCING RELEVANCE
Could not fully rebuild the search elements on site. I worked within the constraint — connected with engineers to visualise the technical flow, then redesigned how diversity appeared without full element refurbishment. Instead of personalisation dominance, I balanced relevance with exploration.
05 · COLLABORATION UNDER PRESSURE
Engineers. Stakeholders. Uncomfortable conversations.
ENGINEERING ALIGNMENT
Engineers raised feasibility concerns on the recommendation card UI — specifically around surfacing personalisation reasoning without exposing proprietary ranking logic. I collaborated directly to visualise the technical flow before designing solutions, not after.
STAKEHOLDER PRESSURE
Stakeholders were pushing for speed. I brought real user feedback — including verbatim quotes like "I feel the system is deciding for me" — into weekly squad workshops. Made it uncomfortable sometimes. Always kept it grounded in what we heard.
CROSS-FUNCTIONAL DOCUMENTATION
Documented every design decision in Notion with explicit reasoning. Not just "we chose X" but "we chose X because users said Y and engineers confirmed Z." Teams could move fast without re-litigating decisions already made.
07 · IMPACT
The numbers are real. Here is what they mean.
TRUST IN AI
Post-test surveys. Users went from skepticism to confidence across buyer and seller groups.
+40
EXPLORATION TIME
Buyers spent more time discovering. Discovery without manipulation — exactly what they asked for.
+26%
LISTING TIME
Sellers completed listings faster. Efficiency without loss of control — exactly what they needed.
−35%
"This feels like Etsy helping me and supporting me sell my products."
SELLER · POST-TEST FEEDBACK · This is the real outcome.
08 · REFLECTION
What I learned about designing trust at scale.
THE ASSUMPTION THAT COST ME
I assumed explaining AI once would be enough. It was not. Users needed ongoing transparency woven into every interaction — not a single moment of clarity at the start. The tooltip was never the solution. The system had to speak for itself, continuously.
WHAT I WOULD DO DIFFERENTLY
Involve the data analyst earlier in defining what "trust" looked like quantitatively. I had the qualitative signal — "I feel the system is deciding for me" — but translating that into a measurable hypothesis took longer than it needed to. Quantitative framing upfront would have accelerated the solution phase.
"The real outcome wasn't the metrics.
It was hearing a seller say —
'This feels like Etsy helping me.'
That's when I knew I got it right."