EM Concepts

Live prototype

Edward Martin sells tiles, lighting, furniture and other home goods online. I shaped EM Concepts to help customers explore how those products could work together in their own space, then develop a room idea they wanted to pursue.

The experience centres on an interior design concept: a room image, a product selection, a title and the reasoning behind the design. People can explore alternatives, choose a direction and refine it while keeping the products connected to the scene.

The problem

A tile, lamp or sofa on a product page is hard to picture in your own room. Styled scenes offer inspiration, but leave customers to imagine how different products would work together.

Edward Martin’s interior designers faced a similar gap. Their paid service started with moodboards because 3D rendering took significant time and budget. Generating room variations could give clients more concrete options to discuss.

The challenge was to support both first-time customers and designers: exploring ideas, inspecting products and developing a direction without losing earlier alternatives.

My role

As Head of Product Design, I owned product definition, the roadmap and the UX architecture. Given its importance as the company’s first AI project, I took a hands-on lead in delivery and designed most of the experience myself. I worked with an AI developer and the CTO, supported by a UI/UX designer on smaller design tasks.

After leaving Edward Martin, I continued the customer experience independently. The Refine workflow described below belongs to that continuation.

How the idea took shape

The CTO and I mapped possible inputs, outputs and actions before and after generation. I turned that into a roadmap ordered by implementation complexity and the inputs each capability needed. The CEO approved the direction.

I used Claude wireframes as starting points for discussions with the CTO and CEO. Those conversations helped establish a compact composer: people could choose a space, style and products visually, then use a prompt to add direction. Further settings opened when needed.

I turned the chosen direction into our first Shape Up pitches, giving design and engineering a defined scope and sequence for the build.

Key decisions

Build the harder flow first

The customer journey had two starting points: a product they wanted to see in their room, or an existing scene they wanted to adapt. Someone planning a renovation could also work from a reference before they had a finished room to photograph.

The interior designers needed a broader workflow. They had to organise concepts into projects, associate them with clients and return to each idea through repeated iterations.

I sequenced that internal workflow first. Starting with the simpler shopping journey risked rebuilding its foundations when project organisation and ongoing iteration were added. We then mapped the customer journey against the shared generative interface.

Give the concept room

The CEO favoured a sidebar with every generation option visible. I argued for a compact composer with controls that opened when needed.

The deciding argument was space. The room image, products and design rationale needed to be considered together. Keeping generation settings compact gave that work more of the screen while leaving detailed controls accessible.

Carry product identity from the input

I proposed using Shop This Look scenes as a library of curated product bundles. EM’s interior designers had already selected products that worked together. Categorising those scenes by style, space and colour could make their judgement part of the generation input.

People would be able to choose a recommended set or select individual products. Product identity would come from those catalogue selections; resemblance in a generated image would not establish which product it represented.

The bundle mechanism was designed and discussed with engineering, but not implemented before I left. Product tagging, including positions within generated images, had been implemented and tested.

Explore alternatives, then refine a direction

In the independent continuation, I developed Refine around a different task from initial generation. Explore lets people try directions through space, style and product inputs. Refine starts with a concept they already like and helps them decide what to change.

I kept scene and style inputs out of Refine’s composer and made the chosen room central. People can select a product through a hotspot in the image or find it in the product list. Both routes open the same detail view, where they can inspect the product and its variants without losing sight of the room.

A variant choice prepares an editable request. A context chip identifies the product being addressed; the prompt describes the proposed change. People can adjust the wording before generating, or use the prompt to request a broader alteration.

This gives a familiar product choice a role in directing the AI. Someone who has already selected what they want to change does not have to describe the whole request from a blank field.

Separate changing a product from buying it

The product someone wants to change and the products they want to buy are different selections.

I kept purchase selection separate from refinement targeting. People can inspect the products associated with the current scene and select all or a subset for purchase, with the corresponding total. Refining the room does not require making a purchase decision, and selecting something to buy does not make it the target of the next change.

Keep each version together

Refinement creates another version within the same concept. Each version keeps its image and supporting context, including the products, variants and price associated with that result.

History switches those together. Returning to an earlier image also restores its product configuration, rather than leaving the person looking at one room with another version’s product list. They can move between results to review the changes without losing later alternatives.

A completed generation becomes the selected version automatically. Choosing a result in History saves that choice too. The concept card in Explore follows the selection, and reopening Refine resumes it with the history intact. Done simply returns to Explore.

What we dropped

We explored a conversational interface for the original product. It worked as a demonstration, but results scrolled away as messages and each change waited its turn in the thread.

A room idea needs to be revisited and developed. I moved the experience into a workspace where concepts remain available and the composer supports the next action. Refine extends that direction by keeping changes within the chosen concept and its history.

How the original work was informed

While the final UI was still being designed, we observed the interior team using an early engineering prototype. The aim was to understand how they wanted to work with AI and how that differed from their rendering workflow. It was an observational session, not a usability test of the final interface.

Once the internal direction was established, we mapped the e-commerce journey against it to check whether the same interface could support a customer arriving from a product page.

The later Refine prototype illustrates the designed interaction model. It does not establish usability findings or production generation fidelity.

Outcome and continuation

At Edward Martin

When work paused during downsizing, the engine generated scenes from scene and style references and could exchange products through prompts, although those exchanges were non-deterministic. The internal tool’s UI covered almost the full journey; the customer-facing journey was less developed.

Product tagging had been implemented and tested. The curated-bundle mechanism remained a design proposal.

Independent continuation

I continued the customer experience after leaving EM. The Refine demonstration now connects product inspection, editable requests, version history and purchase selection within a chosen concept.

The main interaction model is defined. Detailed behaviours for Crop, Select area and Draw still need work, along with production methods for catalogue matching, quantity derivation and authoritative pricing.

The prototype uses prepared scenes and example data to demonstrate the experience. Generated imagery is labelled, with a reminder that colours, shapes and finishes may differ from real products. A catalogue connection gives the concept a product identity and a purchase route; it does not guarantee an exact visual reproduction.