Manufacturer discovery was broken for independent founders.
Fashion entrepreneurs searching for manufacturing partners navigate a landscape that hasn't changed in decades. Alibaba listings with no quality signals. Reddit threads of anecdotes. Spreadsheets traded in private Slack groups. The gap: no platform built for the way small founders actually discover suppliers.
I led this work as one of two founding designers, partnering closely with three engineers. This case study covers the surfaces, research, and systems I drove end to end.
12 founders. The same four frustrations every time.
Semi-structured discovery interviews with early-stage fashion founders across New York and Los Angeles. Three prompts: describe your current process, walk me through your last bad experience, what would make you trust a platform recommendation.
Affinity mapping across 12 sessions. Four friction clusters shaped every design decision.
I know exactly what the product should look like. I just have no idea how to translate that into something a factory in Vietnam will understand.Streetwear founder, NYC · Interview 04
We lost $8,000 to a supplier on Alibaba who ghosted us after the deposit. There was nothing to verify they were legitimate.Sustainable knitwear founder · Interview 07
Research insights: key findings mapped to design implications.
What we got wrong first
The original scope treated this as a search problem: founders type what they want, the NLP model matches it to a supplier catalogue. By week two of interviews it was clear that framing was wrong. Founders weren't struggling to search. They were struggling to translate a vision in their head into terminology a factory would recognize. Search assumes you can name the thing you want. Ours couldn't.
That reframe scrapped the keyword-only concept the engineering team had already started scoping, and pulled the NLP roadmap back to a simpler model paired with image input as a first-class channel, not a fallback. It was a two-designer call. My co-designer and I worked through it together in a working session and brought the reframe to engineering with the interview evidence attached, not just an opinion.
Three principles. Every decision traced back to them.
Meet users where knowledge ends
Accept image uploads and natural language as first-class inputs. No manufacturing terminology required.
Surface trust, don't assume it
Every supplier card shows certifications, MOQ ranges, turnaround windows. AI confidence scores visible and explained.
Make AI legible
Founders see why Pomu recommended a supplier. Every AI output has a UI pattern that makes it comprehensible, not magical.
From intent to matched manufacturer in five steps.
The key structural challenge: progressive disclosure, showing enough to build confidence at each step without overwhelming a founder encountering manufacturing discovery for the first time.
Three IA decisions that shaped this flow
These were worked through by consensus with my co-designer: we'd each bring a version to the table, argue it against the interview data, and merge or kill from there. Engineering weighed in on feasibility before anything got locked.
Snap it. Describe it. Source it.
The AI image-search workflow, shipped as a native iOS experience. Upload a design, let Pomu read it, land on a ranked list of verified factories. Seven screens cover the full arc from upload through empty states to order confirmation. Pomu's founders have since moved on to explore other products, so this reflects the app's last shipped state rather than an active, maintained product.
iOS flow: 7 screens. Upload and describe, AI reads the design, ranked results, match detail, empty state, manufacturer chat, and order confirmation.
Upload and describe
Dual entry: drop an image or type a description. Chips seed common categories so a tap starts a search. A progress ring makes the wait feel like work being done, not waiting.
Ranked results and match detail
Every card leads with a match score, not just a price. The AI confidence score is the hero. A sticky action bar keeps Message and Order Now in reach through the whole scroll.
Edge cases and order
Empty state becomes an action. Manufacturer chat keeps the design as a pinned attachment. Order confirmation surfaces deposit and delivery upfront so expectations are set before production starts.
Animated prototype: the full iOS flow from upload to ranked results.
22 screens before a single pixel of colour.
Desktop and mobile wireframes annotated with structural notes and UX flags. Every surface of the marketing site and marketplace app mapped out with the same content model, so the responsive translation had no surprises.
Desktop wireframes: 11 screens. Landing, services, manufacturer portal, team, sign in, sign up, search results, AI image search, image upload, manufacturer profile, designer profile.
Mobile wireframes: same 11 screens reflowed to 390px. Single-column layouts, sticky primary actions, same content model throughout.
One coherent action, not two inputs bolted together.
Search interface and results: co-located inputs, confidence scoring, transparent match reasoning.
feel human
Live prototype: onboarding through to ranked results with dual-input AI search.
.png)
Mobile experience: fully responsive, WCAG 2.1 AA throughout.
Built for a team that would grow beyond me.
Working with three engineers and one other designer, I built the design system as a shared contract, not just a Figma library but a single source of truth engineers pulled from directly via Dev Mode.
Component library, type scale, design system docs, color palette, and grid system.
One brand. Two completely different audiences.
Pomu had to feel credible to a Brooklyn streetwear founder and professional to a manufacturing director in Hanoi, simultaneously. Every brand decision justified against both.
Brand system: logomark, wordmark, and visual guidelines.
LinkedIn campaign templates: built from brand tokens, publishable without Figma.
What shipped, measured.
40% faster production
Measured against pre-system sprint velocity tracked in team reviews.
350+ followers, zero to launch
Tracked via LinkedIn analytics, correlated with publishing cadence after the template system shipped.
12+ founder interviews
Discovery sessions across New York and Los Angeles that shaped every principle and flow decision.
What zero-to-one taught me.
At zero, there are no existing patterns. Every decision is a bet, and the quality of your research determines how well-placed those bets are. The actual skill is ruthless prioritization: knowing which bet to take next.
Working inside an AI pipeline changed how I think about user mental models. The system's capabilities were fixed. The design was the translation layer. Making AI feel comprehensible, not magical, is now core to how I approach any ML-adjacent product surface.