Agents for Building Design
Adobe has been in digital design since 1982. A now bulky, technical software once transformed analog processes into digital workflow for professionals. It became the standard for multimedia design.
Fast forward to the cloud era, Adobe missed the opportunity to leverage network effects on the web and gave space to Figma, a browser-based and multi-user platform from the ground up that is now valued at $12.5 B.
Now, AI-powered prototyping tools like Lovable, Bolt, and Replit pose a new challenge to Figma. Trained on vast datasets of UI code, they can generate and iterate designs rapidly, replacing PMs with Design Engineers
Zoom out, and the trend is visible in other industries as well
offline tools → cloud-based lightweight applications → Agentic end-to-end automation.
While industries like finance, healthcare, and engineering upgrade tooling, construction design still remains stuck in 2000s with tools like Autodesk Revit.
When working with some friends with background in structural engineering, we interviewed 100+ architects, engineers, and industry thought leaders to understand technological bottlenecks in the design space.
All themes had a common thread: manual file-sharing, slow design, and compliance checks.
There isn't any cloud based tooling because building designs can't be drawn as fast as buttons for home screen while another designer makes the navigation components at the same time. If two people can't make meaningful progress together, it defeats the purpose of rapid, iterative collaboration.
Buildings must have structural integrity. Beams and columns need to be accurately placed for them to withstand load. All of this cognitive work requires calculations.
We figured this to be the perfect application of AI agents. Models that understand the architectural constraints, compliance rules, and modular design principles. If building components can be designed faster, there can be collaborative tools, or even AI-based end-to-end tools for building design.
BuildSync: Agents for 3-D Building Modeling
Inspired by the idea, I worked on BuildSync: an agentic workflow for creating 3-D building designs faster. We built a multi-agent system that leveraged a repository of functions for creating rudimentary components that can be called in different permutations to create different designs. While there are diffusion models that create great visuals, 3-D building models require decimal level accuracy, and therefore, agents that determine parameters for objects are a better fit.
The following diagram illustrates the agentic workflow:
The BuildSync Agentic Pipeline
The BuildSync workflow integrates 23 specialized tools to transform text inputs into accurate 3D building designs, with validation and refinement loops.
The agents make changes at the file, and we use an open source framework for design files, IFC. This ensures interoperability, allowing designs to be exported and modified across different platforms without compatibility issues.
Our work with BuildSync won the Stanford Student Impact award, awarded to students for innovative contributions in AEC (Architecture, Engineering, and Construction) space. To find industry applications, we worked with some of the best architects in the world, including those who designed Burj Khalifa and Saudi's Line Project. We were selected to continue this as part of Autodesk's Research Residency.