The new front door is already in the photo of your own facade
Client
Group A (Kwadro & Belisol)
Group A (Kwadro & Belisol)
Year
2026
2026
Sector
Visual AI
Visual AI
The Challenge
Group A sells windows and doors through Belisol and KwadrO. That happens at the kitchen table, with a sample book and a configurator at hand. You can show the product. Not the house.
That's where the gap is. A customer torn between two front doors is really torn about their own facade. How does that model look there, next to that one window, in that light?
The ask was to take that step out of the customer's head and put it on the screen. A sales rep takes a photo of the house on site and gets back an impression with products from the company's own range.
The bar was high, and rightly so. An image that doesn't look right doesn't sell.
Our Solution
At the kitchen table, the steps stay short. The sales rep signs in, uploads one or more photos of the house and marks which windows and doors need replacing. For each selected door opening, they choose a new door from the range, a colour and a handle or pull bar. Windows add the layout and the glazing bars, as a group for anyone who wants everything the same, or window by window.
Everything there is to choose is an existing product. What the customer sees on the screen, they can also order.
A free prompt field closes the configuration. With it, the sales rep steers the rest of the renovation: a new facade, roof tiles in a different colour, a tree that can go.
That configuration goes to an AI pipeline we built specifically for this range. The result comes back in the tool, passes on to the ERP and can be downloaded by the sales rep.
Photorealism lives in the details nobody notices until they're wrong. We detect windows and doors and keep the AI strictly within those zones, so the rest of the house stays untouched. If a shrub or a car stands in front of the window, a dedicated algorithm handles the visible part correctly instead of simply painting over it. The reflections in the original glass stay too. That's the difference between a render and a photo.
At the kitchen table, the steps stay short. The sales rep signs in, uploads one or more photos of the house and marks which windows and doors need replacing. For each selected door opening, they choose a new door from the range, a colour and a handle or pull bar. Windows add the layout and the glazing bars, as a group for anyone who wants everything the same, or window by window.
Everything there is to choose is an existing product. What the customer sees on the screen, they can also order.
A free prompt field closes the configuration. With it, the sales rep steers the rest of the renovation: a new facade, roof tiles in a different colour, a tree that can go.
That configuration goes to an AI pipeline we built specifically for this range. The result comes back in the tool, passes on to the ERP and can be downloaded by the sales rep.
Photorealism lives in the details nobody notices until they're wrong. We detect windows and doors and keep the AI strictly within those zones, so the rest of the house stays untouched. If a shrub or a car stands in front of the window, a dedicated algorithm handles the visible part correctly instead of simply painting over it. The reflections in the original glass stay too. That's the difference between a render and a photo.
Three things went differently than planned, and all three say something about this kind of work. Two brands on one infrastructure turned out to be an illusion: Belisol and KwadrO each wanted their own environment, and Belisol then split again into NL and BE. We pulled the architecture apart and rebuilt authentication and authorisation per environment. The GPUs we had planned to run serverless were impossible to get because of the global shortage, so we run on dedicated hardware. And the pipeline itself was rebuilt several times, because photorealism on a real facade is not a matter of one good prompt.
That took five people working part-time, from February to September 2026. Tech lead and AI specialist, fullstack, frontend, UX/UI, infra.
Three things went differently than planned, and all three say something about this kind of work. Two brands on one infrastructure turned out to be an illusion: Belisol and KwadrO each wanted their own environment, and Belisol then split again into NL and BE. We pulled the architecture apart and rebuilt authentication and authorisation per environment. The GPUs we had planned to run serverless were impossible to get because of the global shortage, so we run on dedicated hardware. And the pipeline itself was rebuilt several times, because photorealism on a real facade is not a matter of one good prompt.
That took five people working part-time, from February to September 2026. Tech lead and AI specialist, fullstack, frontend, UX/UI, infra.
The Result
Belisol NL went live in early July 2026, Belisol BE and KwadrO in late August. Three environments, one tool, every feature set out at the start delivered.
What customers used to have to picture in their heads is now on the sales rep's screen.
Windows and doors are only one version of this problem. A floor, a tile, wallpaper or paint on a wall, a facing brick, a patio cover: all products you can only really judge once they're in the customer's home. The pipeline we built for Group A is tailored to their range. The approach underneath it is not.
The Result
Belisol NL went live in early July 2026, Belisol BE and KwadrO in late August. Three environments, one tool, every feature set out at the start delivered.
What customers used to have to picture in their heads is now on the sales rep's screen.
Windows and doors are only one version of this problem. A floor, a tile, wallpaper or paint on a wall, a facing brick, a patio cover: all products you can only really judge once they're in the customer's home. The pipeline we built for Group A is tailored to their range. The approach underneath it is not.
Do you sell something people first need to see in their own home?
Other work
