Led design on Poshmark's app redesign — set the discovery thesis, partnered with Naver on the design system, and shipped a 60+ screen migration that lifted Feed orders per DAU by 28% while keeping the core business neutral by design.
The first marketplace to weave AI into both the selling and the buying flow — always led by real user pain, never by "let's use AI." Smart List AI collapses the seller's 15-minute listing flow. Cassie — the AI shopping assistant — takes on the two questions buyers leave the app to answer elsewhere: Size & Fit and Price Confidence. Three projects, one thesis: AI is a tool, the user's problem is the point.
Co-founded a consumer design app and led design from zero. Two pivots, three product shapes (shopping → community → gaming), and the creator loop that scaled DecorMatters to 6M users — through Apple "App of the Day" and a complete iOS + web rebuild.
Interior Atlas
Live · Design guide
If Wind Had Shape
Live · Generative art
If Rain Had a Sound
Live · Generative art + sound
A growing collection of ideas I find interesting enough to build — designed and shipped solo with AI-assisted tools. Some scratch a personal itch, some are taste-tests for new tools, some just stay weekend toys. Each one a way to stay curious.
Framing first — which assumption is cheapest to test, which bet is worth the cost.
A great product gives users a draft they own, not an output they receive.
The harder call only ships when the strategic case is clear and the guardrails hold.
Frameworks that turn one project's lesson into the team's next default.
Weekend prototypes — a plant atlas, a room-style guide. See the Vibe Coding collection.
Furniture rebuilds, gallery walls, a growing houseplant collection — home as a craft testbed.
Home, plants, and slow-living notes on the side — bigger audience than I expected.
Poshmark's first GenAI-powered feature — and the first proof point in our marketplace-first AI thesis: the user's pain is the point; AI is just the tool. We rebuilt the seller's most painful 15-minute flow into a photo-in, ready-to-post-out experience that feels like magic, not autopilot. Same thesis now drives the two buyer-side projects that followed: Size and Fit and Price Confidence.
Every item on Poshmark starts with a listing — but the listing was the tax sellers paid to sell. Title, description, category, sub-category, size, brand, condition, color, style tags, original price, listing price — each one a required field. Publish rate sat at 56%, with the biggest drop-off happening right after a seller entered the camera. Casual sellers and first-timers paid the worst price: they didn't know where to start with "free-form" attributes like Title and Description, and many never finished their first listing.
Early prototypes auto-filled everything silently — and sellers rejected it. They didn't trust a black box to represent their inventory, and they pushed back hardest on the fields that felt personal: original price ("how did you calculate this?"), condition ("this is subjective"), description ("would I write it this way?"). The breakthrough was reframing AI from "it does it for you" to "it drafts, you own it". Trust came from authorship, not accuracy.
The original architecture only worked for the one category the model had been trained on. We rebuilt the flow so sellers could pick the category themselves when ML wasn't confident — auto-detect when we could, gracefully hand off when we couldn't. The same feature now worked for every department in the catalog, not a single slice of it.
Dev testing surfaced a release blocker. We solved it with three design moves: start uploading on photo 1 (not after "Done"), promote Tag to the 2nd photo angle (the ML-richest signal), and add an animated shimmer + rotating progress copy so the wait read as productive, not broken.
One front photo triggers category detection — and the camera surfaces guided photo angles from a 4-master-set system (Tops, Bottoms, Shoes, Bags) mapped across every department, so the next photo is the one ML can actually use. Sellers land on a Listing Preview with Title, Description, Category & attributes drafted; a Configurator below holds the fields ML can't infer (Price, Condition, Size override). Tap any field to override and the preview regenerates inline. Edge cases — blurry photos, ambiguous categories, unsupported departments — gracefully fall back to a "we aren't sure, pick one" prompt instead of an error.
14-day A/B at 11% sizing across iOS, iPad, Android & web. Among sellers who actually entered the Smart List AI flow, 68% reached publish — past the 60% casual-seller goal the project was scoped against. Top-line business metrics (Listers, Sellers, Buyers, Orders, GMV) tracked neutral between control and treatment — exactly the guardrail we'd designed for. No cannibalization. Clear segment win.
The North Star was Listings — the share of seller sessions that result in a published listing. Smart List AI was scoped against a 56% baseline, with a target of 60% for casual sellers and an explicit guardrail to keep top-line metrics neutral.
First shipped GenAI feature at Poshmark. The "draft by AI, owned by seller" framing — and the photo-angle & latency patterns it pioneered — became the team's template for every GenAI feature that followed.
The biggest lesson wasn't about AI — it was about authorship. When we over-automated, we eroded the seller's sense of craft. When we underplayed it, we lost the time savings. The sweet spot was giving sellers the feeling of curation with the speed of automation. The second lesson was about latency as design: a 5–20 second wait reads as broken; the same wait, scaffolded by motion and copy, reads as productive. Both ideas now guide every AI feature my team ships.
Smart List AI proved the thesis on the seller side. Cassie — our AI shopping assistant — extends it to the buyer. Two studies in May 2026 (12 active AI shoppers + 50 Poshmark shoppers) surfaced seven recurring themes; Size & Fit and Price Confidence were the only two that showed up unprompted at every stage of the buyer journey. Those two became Cassie's MVP.
"It's just not knowing what's gonna fit me, so I'm hesitant to buy something." — Poshmark shopper, Study 2
Resale is where sizing anxiety compounds: no returns on many items, no dressing room, brand-to-brand inconsistency, and inventory-of-one. Half of the shoppers we interviewed said they Google items, dig into size charts, and read reviews on other sites before they'll trust a Poshmark listing. Cassie's Size & Fit agent brings that whole loop into the listing itself.
Cassie gives a confidence-scored verdict but never buys for you. Every claim can be interrogated — the "your saved sizes" phrase is a link into the profile that produced it.
The scan runs in phases with named states — "Scanning size charts…" → "Comparing measurements…" → answer. Same pattern I authored on Smart List AI: a 4-second wait reads as work, not lag.
Fit Insights bars (Runs Small 18% · True to Size 73% · Runs Large 0%) show where the recommendation comes from. Shoppers told us they trust an answer more when they can see the crowd behind it.
"Ask anything" appears after the answer. Research showed shoppers want speed first, then dialogue for the edge cases — never a blank chat prompt as step one.
The other need Cassie's research surfaced at every funnel stage. Resale has no standardized pricing, so shoppers do their own comping — jumping to eBay, Google Shopping, brand sites. The Price Confidence agent brings comps, brand-median, condition adjustment, and offer history into one answer, then invites follow-ups ("show me something similar but cheaper," "is $145 in line for this condition?").
No generic AI knows what buyers actually paid for this brand/size/condition across 90M+ listings. That signal — priced offers, sold comps, hold-time — is the honest answer generic AI can't give.
Chip → headline verdict → why → aggregate signal → follow-up. Two agents, one Cassie surface — so the shopper learns the shape of the tool once and reuses it.
Fit unlocks intent; price closes the deal. Shipping Size & Fit as MVP and Price Confidence as fast-follow matches the funnel — and lets each agent's real usage inform the next iteration.
Poshmark is the first marketplace to weave AI into both selling and buying — and always led by real user pain, never by "let's use AI." Smart List AI collapsed the seller's 15-minute listing flow. Cassie takes on the buyer's two hardest questions — will this fit me? and is this a fair price? — the two questions that make shoppers leave our app to answer them elsewhere. Same operating principle throughout: AI is a tool. The user's pain is the point.
I led design for three of Poshmark's highest-intent surfaces — home, search, and filter — through an 8-month system-wide redesign that launched in February 2026. We rebuilt home as a ranked discovery engine, modernized search and mobile filters, and carried a 60+ screen system migration.
The full app redesign wasn't the first thing we proposed. It was the second. The first was a much smaller bet: turn Poshmark's social feed — a busy, follow-based stream — into a personalized, cleaner, ranked feed. I designed and shipped that as a standalone project. It was simpler on the surface and heavier under the hood: ranking, unit types, restraint on chrome.
Simplicity won on the metrics. Engagement moved, buyer intent moved, and — most importantly — nothing degraded on the way there. That result became the case for the bigger bet: if a calmer, more personalized experience worked on one surface, it should work across the app. That's how the full redesign got greenlit — on evidence, not vibes.
Home was optimized for browsing inventory, not deciding what to look at next. Search was a business-critical funnel wrapped in dense, web-shaped filter patterns. The design system couldn't support modernization because no single surface carried enough accountability to force migration. Fixing any one in isolation would miss the real problem: discovery needed a single spine.
Before rebuilding, we audited the existing app against buyer intent and comparable consumer benchmarks. Four issues kept surfacing — each one solvable in isolation, but together they explained why the app felt dated and why younger audiences bounced.
Typical redesigns get judged by top-line engagement and quietly fail — traffic moves around inside the app, and nobody can tell signal from regression. We designed against an explicit mix-shift hypothesis: attention should flow into Feed and Search, out of Brand and Community Closet, and OI / DAU should stay neutral overall. Naming the expected losses upfront let us defend the wins.
Moderated user-call transcripts (Pinetail Research × Poshmark, 8 participants) clustered into four pillars. Together they describe how it felt to navigate the new app — not what it did, but what kind of product it felt like.
"I love the new look and feel. It's very clean, organized, and easy to navigate." — P01, moderated user call
Before pixels, we agreed on directional intent. Four axes, each marked from where the app was to where it needed to be. This map became the tie-breaker whenever exploration spread too wide — and made it explicit that we weren't flipping Poshmark's identity, we were re-centering it.
Legacy product photos were square, which suited general marketplace inventory but not apparel. Moving every product card to a 4:5 vertical ratio aligned photography with how sellers actually shoot outfits, tightened rhythm across the grid, and gave UGC room to breathe without cropping heads or shoes.
Home stopped being a product wall. It became a ranked, modular feed that mixes personalized product grids with UGC photos and editorial lookbooks — a reason to open the app when you don't already know what you're looking for.
Recommend highly relevant products based on each shopper's interests and activity history.
Vertical photo ratio, optimized for apparel — full outfit shots, not cropped squares.
Auto-scrolling carousels curated by theme, giving the feed a rhythm beyond product tiles.
The old search stacked filters, chips, ad labels, and a 2-column grid into a single dense screen. The new surface breathes: search moves to its own bottom-nav tab, filter chips become horizontally scrollable pills, product cards lead with imagery, and meta drops to a second line. Home became a ranked feed of modular units — product, creator, live, show — and mobile filters collapsed into a flatter bottom-sheet IA.
For shoppers who could describe a vibe but not a query, Posh Lens turned the camera into the search bar — point, snap, get back ranked results from the catalog. Built on the same ranking spine as the redesigned Search tab, so visual queries inherit the same filter chips, ad logic, and feed treatments shoppers already learned everywhere else. Same mental model, new entry point.
Page View / DAU mix-shift. Neon = intentional gains. Dashed = expected losses from removing the Shop tab.
| Surface | Pre | Post | Δ |
|---|---|---|---|
| Feed | 5.6% | 7.2% | +1.6 pp |
| Search | 44.3% | 45.4% | +1.2 pp |
| Brand | 10.9% | 9.7% | −1.2 pp |
| Show | 15.3% | 14.6% | −0.7 pp |
Feed's +9% OI/FM was the bigger signal than traffic: buyers didn't just visit more, they bought more of what they saw. New-user metrics moved up across the board — D2 retention +15%, D1 sessions +10%, D1 buyer +9%. Ad revenue still grew +2.6% overall despite Brand losing top-of-funnel.
"The new app is much better — the redesign makes it easier and the flow works so much better. 10/10."
"I love the new look and feel. It's very clean, organized, and easy to navigate."
"The larger photos make listings look great and really improve the presentation."
"It's much more intuitive and sleek than the previous version."
"The new seller experience is great! I love the new app design."
"Overall, I like the new app better. It feels cleaner and more modern."
We couldn't run a clean A/B (the redesign shipped 100% at launch), so credibility came from repeatedly cutting the data — cohorts, channels, tenure, platform — and showing the same directional lift each time. Pre period fixed at Mar 13–19, 2026 (old); post window Apr 1–7 (new). All Δ below are statistically significant at p<0.05 unless noted.
This project reset how I think about product design. The job wasn't to "improve home" or "redesign search" in isolation — it was to decide where attention should flow across the marketplace, and to defend that decision with a measurement framework explicit enough to tell mix-shift apart from regression. Design systems rarely win on abstraction alone; attaching the 60+ screen migration to surfaces with real traffic made it impossible to ignore. And the story arc — ship the small proof, use it to earn the bigger bet — is how I now approach every ambiguous redesign.
I co-founded DecorMatters and led design from zero. We shipped, we were wrong, we pivoted — twice. Each pivot was a design problem I ran end-to-end: insight, research, system, launch. This is the story of how the product became what it was always trying to be.
DecorMatters shipped three times. Each version answered a different question, each pivot was a design call I made with the data in hand. Here's the timeline — then three chapters, one per pivot.
2017. AR on iOS was brand new. Our thesis: people don't buy furniture online because they can't tell if it fits. Drop a real chair into a real room, hit buy. Simple.
I don't come here to shop. I come here to design my dream rooms when I'm bored.
The returning user was 25–55, mostly female, coming back 3–5 times a week to create, not buy. We didn't need a better catalog. We needed a design tool, a social feed, and something to do with free time. Pivot.
After the first pivot, design was the loop. But without publishing there was no feed, no likes, no reason to return. The number that mattered was stuck at 3.5%.
I led design on three parallel tracks — speed up item search, make tools forgiving, give publishing a reason. Each shipped as a minor release so we could measure independently.
By 2020, MyDecor subscription drove ~75% of revenue — fragile. The goal was 2× without breaking the free creator experience. The framing shift: stop treating this as retail. Start treating it as a game economy.
Most comments were "beautiful / wow / awesome." We turned the compliment itself into the product — paid virtual gifts attached to designs you loved.
Daily check-ins earned Dcoins; finished challenges won badges. Gamers' real needs — practice, recognition, a fun escape — became the loop that kept them opening the app.
Instead of a hard paywall, items could be unlocked three ways: membership, purchased Dcoins, or earned Dcoins. Free users kept designing. Invested users paid — or earned by showing up.
Five years of DecorMatters compressed into the lessons I actually carry.
A small, growing collection of AI-built side projects — designed, written, and shipped solo with Claude. Part sketchbook, part proof-of-craft. Each one a small bet on a tool I want to understand by actually shipping with it.
I started this collection to stay honest about what AI is good at — and what it isn't. Every project here is a hypothesis about a tool, an interaction, or a product. Some inform my work. Most teach me something I couldn't have learned from a blog post.
A bilingual (EN / 中文) reference site documenting 26 interior design styles — their origins, defining traits, materials, color palettes, and the moods they make. Built with Claude and Next.js over a couple of weekends.
Spin the globe, discover the world's most extraordinary plants — baobabs, dragon's blood trees, welwitschia, and others mapped to where they grow. A 3D WebGL botanical archive built with globe.gl.
A generative p5.js study — first in a series on nature's emotions. A field of reed grass gives the invisible away: breezes comb through it, gusts sweep past in slow soft waves, seed-fluff catches the light. Move your mouse to stir the air; click to summon a gust.
Second study in the nature's emotions series. The rain itself never appears — you only hear it, and watch the ripples it leaves on a jade-green pond. The water is a real-time wave simulation; every sound is synthesized live with Web Audio, from the hiss of the downpour to each drop's plink. Click to listen; move your mouse to stir the water.
Vibe coding with AI changed how I design at work. Shipping these projects solo made me sharper at scoping, faster at prompting, and more opinionated about when AI should disappear into a product versus announce itself. The best ones usually disappear. More to come — this collection grows whenever I have a weekend and a curiosity.