From Inspiration to Reality: Building an AI Home Design System

For people looking to refresh their homes, generative AI offers an enticing possibility: Upload a photograph of a room and, within seconds, see what it might look like with a different color palette, new furniture or an entirely new style.

The results can be impressive. They can also present an interesting problem. The room that comes back may have a larger window, a different fireplace or an entirely new layout. Somewhere in the redesign, the space has stopped looking like the one you actually live in.

For MS in Applied Data Science students Aigul Azamat and Akshat Gupta, that disconnect became the basis for a capstone project, and ultimately, a Best in Show winner at the Summer 2026 Capstone Showcase. The AI home design system is one of the projects developed through the MS in Applied Data Science Capstone experience, where student teams spend two academic quarters applying their skills to real-world challenges through industry-sponsored, research-focused and student-proposed projects.

“Quite often people have trouble describing the room they want, but they recognize it the moment they see it,” Azamat said. “Our app shows them what a professional designer actually did to a room similar to theirs, using real renovations from HGTV’s Property Brothers show, and then restyle their own room to match the one they pick.”

The idea did not begin with interior design. Gupta and Azamat were initially interested in recommender systems and what they saw as a limitation in the way platforms like Netflix suggest what a user might want next. Increasingly, Gupta said, they found themselves turning instead to generative AI for recommendations.

Through guidance from their faculty advisor, Nick Kadochnikov, that broader idea eventually found a more visual application: home design.

Designing Without the Perfect Prompt

Generative AI tools often depend on a user’s ability to articulate what they want. Interior design can make that particularly difficult.

A homeowner may love the warmth of one room, the furniture in another or the overall feeling of a professionally designed space without knowing whether to call the style transitional, contemporary, Scandinavian or something else entirely.

Azamat and Gupta designed their system so users would not necessarily have to know.

A user begins by uploading a photograph of their room. The system searches a database of professionally renovated spaces and surfaces three designs that are similar to the uploaded room. The user selects the one that best reflects what they like.

The system can also learn from a conversation with the user, combining those preferences with the original room and selected inspiration image.

“This tool is very helpful for people who can’t type a specific style into the prompt box because they simply don’t know how to put that into words,” Azamat said. “So instead, they upload one photo of their room. We show them three real finished rooms. They pick one, and they see their own room in the style they chose.” 

And the process does not end with the first generated image. Users can continue modifying individual elements, removing a sofa, changing the color of a rug or repainting the walls, while the system builds upon the previous version rather than starting over. 

Teaching AI What a Renovation Actually Looks Like

Building that experience required more than connecting an existing image generator to a chat interface.

The students assembled a dataset based on real professional renovations, including before-and-after images from Property Brothers. They developed an agentic web scraper that could collect relevant images from selected websites, then paired the original rooms with their professionally renovated versions and textual descriptions of the transformations.

“For every pair in the data set, we keep the room type together with a textual description of the resulting style,” Azamat said. “That gives us three things: the original room, the version the TV show home designers actually built, and a description of their transformation.”

Behind the interface, the students built an agentic system organized around a central planner that coordinates specialized agents and tools responsible for different parts of the process. Once a user chooses an inspiration image, a transformer model trained by the students combines the conversation, inspiration image and photograph of the user’s existing room to generate the redesign.

But generating an appealing room was only part of the challenge.

The harder problem was making sure it remained the same room.

Keeping the Windows Where They Belong

Generative image models can be remarkably good at creating beautiful interiors. They can also be remarkably willing to renovate things the user never asked them to touch.

For Gupta and Azamat, preserving the underlying geometry of a room, its windows, fireplace and structural elements, became one of the hardest parts of the project.

“That is the most technically challenging part,” Gupta said. “You might have noticed that what these tools tend to do is they completely alter the room.”

During training, the students introduced guardrails that rewarded the model for preserving the original room and penalized it when a generated design changed the space too dramatically.

Solving that problem also led to another challenge: determining whether the model was actually producing good results.

The students drew on concepts from courses including Advanced Computer Vision, Statistical Models and Causal Models for Data Science to evaluate their generated images.

“Another challenging part of the project was evaluating the quality of what we generated,” Gupta said. “How we actually evaluate the generated images is with various statistical techniques. That’s where the concepts taught in those classes came up.”

Making Professional Design More Accessible

For Azamat and Gupta, the project also raised a more practical question: Who might actually use something like this?

Their answer was homeowners preparing to renovate, particularly those who know what they like but cannot necessarily translate that preference into the language of interior design or an effective AI prompt.

Professional help can also be expensive. The students’ research found that hiring a professional interior designer can start around $8,000, with costs varying based on the size, scope, and location of the project.

“We did some research, and we found out that most people would probably love to work with professional designers, but not everyone can afford it. ” Azamat said.

The goal was to make elements of the design process more accessible to people who might otherwise go without that expertise.

For now, the project has already gone further than either student expected.

When the Summer 2026 Capstone Showcase concluded, their AI home design system was selected as Best in Show.

“It was very surreal. Honestly, we were not expecting to win,” Gupta said. “This came completely out of the box.”

Azamat remembers seeing the email and immediately messaging her teammate.

“I didn’t expect that at all,” Azamat said. “I cannot believe it.”

From the Classroom to the Capstone

For Azamat and Gupta, what began as an exploration of recommender systems ultimately became a new way to approach challenges in home design. People may not always have the words to describe the room they want. Sometimes, they just need to see it.

Explore more work from MS-ADS students in the Capstone Project Archive. Projects span areas ranging from healthcare and finance to computer vision, robotics, generative AI and more, giving students an opportunity to bring together the technical and problem-solving skills developed throughout the program.

Explore the Project

Want to take a closer look at how Azamat and Gupta built their AI home design system? Explore the project on GitHub here to see technical details and more about how the system works.