George Zheng spent years designing buildings at a major U.S. architecture firm. Now he’s betting that the same training can help millions of real estate agents understand and price the spaces they sell.
Zheng founded Edensign, an artificial intelligence startup incubated at Harvard Innovation Labs’ Launch Lab X program. It turns photos of empty or cluttered rooms into staged, listing-ready images in about 15 seconds.
But Zheng says virtual staging is only the wedge. The company has expanded into what it calls “listing intelligence”: room-by-room condition scoring, improvement recommendations and pricing guidance. It’s built on a spatial model trained to leave a home’s architectural elements untouched.
Zheng spoke with Inman about why he chose real estate agents over architects, how Edensign handles the growing compliance scrutiny around AI-altered listing photos, and his 10-year vision for AI across the entire “space-making lifecycle.”
Inman: Tell me more about Edensign. How did it start and what was your vision for it?
George Zheng: I founded Edensign last year. We were selected for the Harvard Innovation Labs — the Launch Lab X program — which picks about 20 founder companies each year from hundreds of applicants in the Harvard community, so it’s competitive, and we got in. They provided us with office space, and we started with virtual staging as an entry point.
I have an architecture background. I graduated from Harvard’s Graduate School of Design as a trained architect, and before that I worked for Perkins&Will for several years. I saw a problem in the industry, so I started this company.
I think our special angle is that we started with virtual staging and expanded into listing intelligence — market positioning, buyer appeal and so on. I see the industry’s problem as fragmented listing preparation, and most generative AI still treats every photograph as an isolated image.
We found that single-image generation isn’t enough, so we built a spatial layer that supports multi-angle listings — including condition assessment, recommendations and ROI — to help agents make better decisions. That’s what we’ve evolved into.
What convinced you that virtual staging was the problem you wanted to work on?
Virtual staging is only a wedge market. We don’t define ourselves as a virtual staging company today.
I didn’t leave architecture because I stopped caring about buildings. I started Edensign because I wanted to apply my architectural training at a scale traditional practice could never reach.
An architect might influence dozens of spaces over a career. Software and AI can help millions of people understand space and make better decisions. That’s the origin story. I saw the opportunity in the AI shift to apply architectural intelligence to how millions of people understand space.
How do you think coming from architecture shaped the way you built the company?
George Zheng
Architecture teaches you to understand space as a system — geometry, circulation, scale, light, material and how people actually experience it. That level of architectural rigor and domain expertise hadn’t been applied to the real estate agent side yet, and I think that’s the huge opportunity.
I didn’t build a 3D modeling tool for architects, because I think that has less impact. Architecture school teaches you about the top 1 percent of iconic buildings — starchitects, modernism, the flashy stuff that drives academic discourse.
But the other 99 percent of the built environment is generic housing stock — colonial houses, tract homes — that academia largely ignores. Improving that generic built environment by even 10 percent is, to me, more impactful than designing one iconic building. That’s why I chose to build a space-making tool for the 2 million to 4 million real estate agents out there, rather than starting with homeowners or architects.
You mentioned Edensign does more than virtual staging. What else are you working on?
Our differentiation isn’t just what the foundational large models do out of the box. We’re building a domain-specific layer on top: spatial reconstruction, scale and camera-relationship modeling, consistent object placement, and real-estate-specific training and evaluation data. We train our own model on a proprietary dataset, with workflows that tie the visual output back to real spatial data.
Today, an agent can upload one address and a folder of raw photographs, and the platform will identify room groupings, evaluate quality and condition, recommend high-impact improvements and staging choices, analyze comparable properties and marketing positioning, and suggest what drives the most ROI.
Because it’s grounded in architectural domain knowledge rather than generic image generation, the platform maintains architectural integrity while adding more advanced capabilities, including 3D, to empower agents.
Is there anything specific about AI right now that makes what you’re doing possible in a way it wouldn’t have been a few years ago?
I think about it this way: Three years ago, when I was at Perkins&Will, we had a project backed by Blackstone. I designed an office space from scratch through to the final model. Phase one got built, and then we moved into phase two, and because of COVID, no one was renting office space, so phase two basically got lost.
Even a top firm can lose a project like that. I realized decision-making moves so slowly because it’s manual and lacks an objective basis. Developers talk in their language, architects make judgment calls based on gut feeling about whether something looks good, and developers care about numbers and rent. Everyone’s speaking a different language.
AI can now translate that into a more objective, logical decision-making process. We’re not making final decisions for decision-makers, but we provide insights. We de-risk the choice. Here’s option one, option two, here’s the risk behind each, and here’s what you can achieve. That’s the architecture-to-programming-language translation that gets everyone on the same page about the same objective.
The other piece is 3D — geometric precision that lets us deliver at a level and consistency that wasn’t possible two years ago. And from an industry-validation standpoint: Zillow acquired Aryeo in 2023, and CoStar acquired Matterport last year.
Everyone’s buying point solutions, but we want to become the infrastructure for all of that decision-making. The industry validation means more capital is flowing into this space, which gives the technology more credibility and lets a small, early-stage startup like us scale faster because more resources are available.
There’s a lot of talk right now about AI-generated listing photos and growing compliance concerns. Does that intersect with what you’re doing? Are there regulatory concerns around using AI for virtual staging?
That’s a great question, and it reflects one of our competitive advantages, because we train our model on architectural domain-specific data. Our differentiator is maintaining architectural consistency and integrity with high precision.
A lot of general-purpose models, if you tell them “stage this with modern living room furniture,” will also change the lighting, the window size, the sense of depth, the floor material. The space ends up looking different, maybe even larger. That counts as misrepresentation under MLS regulations.
What we do — and I’m an architect, so I built this in — is recognize the house’s current condition and make sure the model doesn’t touch any architectural elements. We maintain architectural integrity and only stage what’s necessary: the furniture inside the space. We call that domain-specific virtual staging, unlike many cheaper AI tools that don’t make that distinction.
On top of that, at a conceptual level, we can tell an agent: based on our analysis of your listing, here’s the score and condition for each room, and here’s what would improve it. For example, we had a client in Boston where we suggested a $30,000 kitchen update — replacing old black cabinets with white ones — that could add $100,000 to the sale price.
That’s the kind of professional guidance we can give sellers and agents. But for anything at that conceptual level, we disclaim clearly: This is a visualization based on our analysis, meant to give you an idea. It’s not the final result of actual construction. We provide these insights as a reference so agents and sellers can make a smarter decision on top of it, and that’s the guardrail for what we present versus what’s presented as a concept.
What’s your long-term vision for Edensign?
Today, the visible output is listing preparation. But the underlying, longer-term question is much bigger: Can AI understand the current condition and potential of a physical space well enough to support decisions across the entire space-making lifecycle, such as marketing and purchasing, renovating, leasing, even real estate development? That turns Edensign from a staging startup into a broader industry thesis about the whole space-making lifecycle.
Ten years from now, picture every space running through an “Edensign” process: what’s the optimal outcome, what’s the goal, and we’ve analyzed all the details. Here’s what to improve, here’s the price, here are comparable sales, here’s the risk — and we hand that to the decision-maker to make a better judgment.
This October, Inman turns its focus to AI and its rapid rise in real estate. During Artificial Intelligence Month, we’re digging into the startups shaking things up and the established players folding AI into their offerings.
