An AI cabinet visualiser that lifted leads 20% for a national online retailer

Context

The client is a national US online retailer of ready-to-assemble kitchen cabinets, running a long-established nopCommerce platform. KodeLogiK has looked after that platform for several years as a dedicated engineering resource: shipping-API migrations, email-marketing integrations, database optimisation and the day-to-day work that keeps a high-traffic store fast through sale campaigns.

The visualiser was a feature inside that relationship, not a separate project. That mattered for how it was built.

The problem

Cabinet doors are a considered purchase. Shoppers compare door styles and finishes for weeks, and the question that stalls most of them is simple: what would this look like in my kitchen? Sample doors help, but they are slow to ship and show one style at a time.

The retailer wanted shoppers to answer that question on the product page, from a photo of their own kitchen, without leaving the store or installing anything.

Our approach

We treated it as a storefront feature first and an AI feature second. The visualiser had to live on the existing nopCommerce product pages, respect the catalogue data already in SQL Server, and degrade gracefully if the model was slow or unavailable.

• Model choice. We evaluated image-generation options against one test: does the rendered kitchen still look like the customer’s kitchen, with only the doors changed? OpenAI GPT-4o’s image capabilities gave the most faithful results with the least prompt engineering, so we built on it.
• Prompt and guardrails. Each door style maps to a structured description (profile, finish, colour, hardware) stored alongside the product. The prompt is assembled from that data, not written per request, so new styles need no engineering work.
• Integration, not a widget. The upload, generation and result views are nopCommerce plugin components, styled as part of the store. Generated images are stored in Azure and tied to the shopper’s session so they can compare several styles side by side.
• Cost and latency control. Requests are queued, rate-limited and cached by image-and-style, so the same kitchen rendered in a second style reuses the uploaded photo and keeps the per-render cost predictable.

What we built

• A photo upload and generation flow on the product page, on desktop and mobile.
• A GPT-4o rendering pipeline that keeps the customer’s room and replaces only the cabinet doors with the selected style.
• Azure storage and a lightweight job queue for renders, with caching per kitchen photo and style.
• A lead capture step: shoppers can save or email their visualisation, which hands a warm lead to the sales team.
• Analytics on which styles are visualised and which visualisations lead to orders.

The result

Shoppers now see their own kitchen in the door style they are considering, and more of them get in touch and order: leads rose 20% and conversion to orders 5%, by the retailer’s own measurement. The feature shipped in three months with two engineers, alongside the usual platform work, and the same engagement continues to keep the store fast through peak campaigns.