Upload a photo of a customer's building, render our product on it

Customers can't picture the product on their own house

A homeowner looking at our product on a website or a sample board still can't quite see it on their own house. 'Use your imagination' is where a lot of deals quietly stall. People want to see the actual thing on their actual home before they commit, and until recently the only way to give them that was to make it by hand.

Reps were burning time on one-off mockups

So our reps did make them by hand, one at a time - a designer photoshopping the product onto a customer's photo whenever a deal needed a nudge. It worked, but it was slow and it didn't scale, and it only happened for the deals someone remembered to chase. The moment a picture would have closed the gap was usually the moment nobody had time to make one.

Upload a photo, see it on your real home in seconds

So we built a tool that does it in seconds. A homeowner uploads a phone photo, marks the opening, picks the product and a color, and gets a preview of it on their real house - no login, no waiting on the office. What used to be a designer's afternoon is now something a customer can do themselves while they are still interested.

A second AI grades every image, so a wrong preview never gets to the customer

The risk with AI images is not that they look bad. It is that they look great and are wrong - the product in the wrong spot, the wrong size, or extra parts that don't exist. So every render is checked by a second AI that grades it against a detailed rubric, re-rolls once if it fails, and attaches an honest caveat if it still isn't sure. We even had to upgrade the grader's brain partway through, because the cheaper model kept waving through the subtle mistakes.

The cheap model kept missing the subtle mistakes, so we gave the grader a smarter brain.

We picked the model with a $10 experiment

Before we wrote a line of the app, we settled the biggest question - which image model to use - with a small, cheap experiment: a fixed set of hard photos, run through the candidates, scored the same way. It cost about ten dollars, and it saved us from building the whole thing around the wrong engine. Decide the expensive question with data, not a hunch.

We spent about ten dollars settling which model to use before we wrote a line of the app.

The AI drew audio cassette tapes instead of curtain cassettes on a house

Not that it went smoothly. We call the head-box of our product a cassette, and the model took us at our word - it rendered literal audio cassette tapes, spools and all, on the side of a house. It was a perfect little lesson: your industry's words are not the model's words. The fix was a growing library of very specific instructions about what the product is and, just as important, what it is not.

We call the head-box of our product a "cassette." The AI took us literally and drew audio cassette tapes, spools and all, on the side of a house.

A convincing-but-wrong preview is worse than nothing

That is the rule the whole tool is built around, and it is the one we would pass on: with customer-facing AI, a confident wrong answer costs you more than no answer. It is why the grader exists, why we ship an honest caveat instead of a slick lie, and why this is still something our reps hand to specific customers rather than a button we have turned loose on the whole internet. Get the trust right first. Then scale it.

A convincing-but-wrong preview is worse than no preview at all.