Interior Trends:
AI-Powered Discovery
for the German Furniture Market

A fragmented market where buyers can see what they want but can't find it

The German furniture and home goods market is one of the largest in Europe — and one of the most fragmented. 200+ active retailers. No unified product taxonomy. The buyer journey looks like this: see a sofa on Instagram, save the screenshot, spend twenty minutes searching for "grey velvet sofa" across Westwing, Made.de, Otto Home, and a dozen others, give up.

Existing search tooling is text-based and category-bound. You can filter by brand, price, size. You can't search by what something looks like, by what fits the room you already have, or by what's available across retailers.

Interior Trends starts from a different premise: search should be visual, contextual, and cross-retailer. Type less. Upload more. Compare automatically.

0→1 with no map

• No reference pattern — no established image-first furniture discovery product to learn from, in Germany or elsewhere.
• No existing research — zero prior user data; every assumption about how people shop for interiors had to be built and tested from scratch.
• A small team and budget — a single engineer on the build; every design decision had to be cheap to build or clearly worth its cost.
• A fragmented market — 200+ retailers with no unified product taxonomy to search across.

End-to-end product ownership

I'm the founding product designer and equity partner, working alongside the technical founder. I co-manage the team with the CEO, and my scope covers product, design, and research — full ownership of what gets built and how it ships, including investor conversations.

• Product strategy — roadmap, feature prioritization, monetization model design
• User research — 15+ interviews and usability sessions with German shoppers, ongoing
• Product design — every screen, flow, and interaction across the platform
• Design system — token-based foundation built for rapid iteration
• Marketing collaboration — landing pages, AI feature positioning, content design

This is the highest-leverage track in my work right now. Full ownership, an AI-native product live in market, no compliance friction blocking publication. Every shipped feature is a publishable proof point.

A diary study set the direction — before a single screen

I started with an 8-person diary study with German homeowners over two weeks. The finding reframed everything: interior discovery is visual and mood-driven, not keyword-driven. People react to a room, not a spec sheet — nobody types "mid-century walnut sideboard, 120cm." They think in moods and spaces: "something that feels like a Scandinavian cabin but fits a 12m² room."

That one insight defined the AI interaction model — similarity and room context over text search. It was the most valuable two weeks of the project.

From there the path was deliberately short: interaction model → Figma → a clickable Framer prototype → usability sessions (15+ to date) → a six-week MVP build with one engineer. The interaction model was validated with real users before hand-off, so the engineering build went into something already tested, not a guess.

The search box was the wrong starting point

The first prototype had a search box at the top. Users typed "scandinavian dining table" and got results. It worked. We almost shipped it.

The user research killed it. Across 15 interviews, the same pattern: shoppers don't have words for what they want. They have a screenshot from Pinterest, a photo of their living room, a reference image saved from a friend's house. Forcing them to translate that into a text query is the friction we were supposed to remove.

We rebuilt the discovery surface around image upload. The user drops a photo — a screenshot, a sketch, a corner of their actual room — and the system returns visually similar products from connected retailers, with cross-retailer price comparison.

Text search stays as a secondary path, not the front door.

The hard part of AI-native design is trust, not generation

Generating recommendations is the easy half. The design problem is getting people to trust and act on them — a surface that reads as considered, not as a black box. I designed for that on three fronts:

• Natural-language contextual search — describe the feeling, not the SKU.
• Image-based exploration — upload a room photo, find products that fit it.
• Recommendations that explain themselves — enough "why this" that a result reads as considered, without burying the user in controls.

The design discipline was restraint: expose just enough of the model's reasoning to build confidence, and hide the rest.

AI as the spine of the product, not a sticker on it

Every feature below was scoped against one question: does this make image-first discovery work better, or is it a feature for the sake of a feature?

• Image-Based Product Search — upload a photo, get visually similar products across all connected retailers. Built on category-specific embeddings tuned for furniture aesthetics.
• Contextual Recommendations — AI considers browsing history and previously uploaded room photos to surface products that fit the user's emerging style profile.
• Cross-Retailer Price Comparison — same or similar product across multiple stores, ranked by price, availability, and shipping time.
• Saved Boards / Wishlist — collections that double as input for the AI recommendation engine; the more a user saves, the sharper the suggestions.
• Landing Pages and Brand System — designed alongside marketing collaboration to land organic traffic and convert image-search visitors.

Token-based foundation built for a small team shipping fast

The team is small. The system has to do the work of three designers.

• Atomic token foundation — colors, typography, spacing, elevation
• Component library covering 80%+ of UI surfaces by month two
• System lives in code — tokens and a Storybook-documented component library for clean frontend handoff
• Built to support light and dark themes from day one

The system is the reason a single product designer can ship a complete e-commerce platform with AI features without slowing the engineering team down.

Shipped, live, instrumented

• Live in the German market — the product is launched and iterating with its first users
• 0→1 product built from scratch — research, design, system, shipped
• Design system in code — tokens and Storybook-documented components for clean frontend handoff
• 15+ usability sessions completed, validating the image-first approach across multiple shopper segments
• PostHog event tracking on every key interaction from day one — iteration runs on real behavior, not opinions
• Founding scope — equity partner co-managing the team with the CEO: prioritization, roadmap, investor conversations

AI as product spine, not feature sticker

Most "AI-powered" e-commerce features are bolted on — a chatbot in the corner, a recommendation row tagged "AI Picks," a generated description block. They don't change the underlying flow. Users don't notice them. Adoption stays under 5%.

Interior Trends starts from the opposite assumption: AI is the discovery surface itself. Remove image search and the product collapses. We don't push users into the feature with onboarding nudges — the feature is how the product works, and that's exactly what we're now measuring with our first users in PostHog.

This case is the proof that AI in product is a positioning question, not an integration question. You either build the product around the AI capability and earn adoption, or you ship a feature that nobody opens twice.

0→1 needs more research, not less

The temptation in a 0→1 build is to move fast and skip the research. The diary study was the opposite call — and it's the only reason the AI model matched how people actually shop instead of how we assumed they would. The search box felt obviously right; the users proved it obviously wrong. I'd make that trade — two weeks of research against weeks of building the wrong thing — every time.