Steel kitchens are robust, durable – and often modular. This is exactly what makes them ideal for resale. But which modules can be reused? Our AI-supported proof of concept for Forster Küchen shows how potential can be quickly identified and exploited using smart tools.

With an extensive catalog containing several hundred elements in five design and size variants, manual identification can quickly become a challenge.

This is exactly where our approach comes in: a responsive web app that uses AI-supported image recognition to enable structured recording of the installed elements with minimal effort by the user.

Our goal: to provide targeted support for a typically manual process through automated AI image recognition – and thus make a contribution to the circular economy in the kitchen sector.

Own image recognition vs. existing AI models

An initial approach was to train our own image recognition model – technically fascinating, but time-consuming and ultimately superfluous: What we needed already existed. GPT-4o from OpenAI has exactly the capabilities that were crucial for our goals.

Limits and challenges

Nevertheless, the implementation remained challenging. One of the biggest technical hurdles was the precise interpretation of the images. So we asked ourselves the question: What features can multimodal models such as OpenAI’s GPT-4o recognize in our images of kitchen elements? Shapes and objects can be recognized from a two-dimensional image, but exact proportions? Not a chance.

Another area of tension was prompt engineering. Initial attempts to include all information and requirements in a single prompt led to inaccurate or incomplete results. Although the models recognized the picture, they overlooked important details or interpreted them incorrectly. Reliable functionality could only be achieved through iterative optimization – i.e. splitting into smaller, targeted prompts.

The centerpiece is a Next.js application, combined with the Vercel AI SDK and the new GPT-4o model from OpenAI.

The real “aha” moment came when we fleshed out the idea of a “wizard” for this case – a guided assistant within the application that systematically guides users through a filtering process. The aim was to gradually retrieve relevant information up to the kitchen element being searched for. In the next step, we combined this wizard with intelligent image recognition: uploaded kitchen images were interpreted by the AI so that kitchen elements could be roughly recognized and many of the questions answered automatically. This made it possible to minimize user interaction.