Interview: Parafin Founder Adam Hengels
Interview: Parafin Founder Adam Hengels
Architecture, machine learning, and the future of designing everything
Generative AI is all the rage. And while tools like ChatGPT and Midjourney are revolutionizing content creation, similar approaches are being applied to architecture. But instead of taking inputs like Harry Potter x Balenciaga, tools like Parafin let hotel developers specify site conditions, design preferences, and financial outcomes to get a design back in seconds (instead of days or weeks).
The following is an interview with Parafin co-founder Adam Hengels. We cover his road to starting the company, how machine learning (ML) tools change the way architects work, and the broader opportunities for applying technology to real estate development.
Introducing Adam
Jeff Fong: Adam, thanks for sitting down with me. Walk us through your professional journey leading up to Parafin.
Adam Hengels: My pleasure, Jeff. In high school, I became fascinated with CAD drafting. I initially thought that being a CAD draftsman was my calling, but I was advised that it might not be the most lucrative career choice. So, I decided to pursue architecture. However, after two years in architecture school where we weren't allowed to use computers, I switched to engineering.
During my studies, I worked at an engineering firm as a CAD draftsman and later practiced structural engineering. But my entrepreneurial spirit led me back to school to earn a master's degree in real estate development from MIT. After that, I spent about eight years at Forest City Ratner, working on the Atlantic Yards and the Barclays Center project. Eventually, I started my own development company focusing on innovative housing types for young people, such as micro housing and shared living.
After returning to Chicago from New York, I was introduced to my co-founder, who had a background in architecture with a focus on computational design. Combining my developer mindset with his computational design expertise, we began exploring ways to optimize building designs using algorithms. While optimizing window placement for daylighting and energy efficiency was impressive, I suggested we focus on optimizing the overall building design for profitability.
JF: Is it fair to say that your co-founder had the list of solutions and you had the list of problems (and knew which ones were most worth solving)?
AH: Yeah. He had the techniques and I had the problems that could be solved with those techniques.
Parafin the Product
JF: So that's a good jumping off point directly into Parafin. It sounds like there's a design piece and a financial optimization piece, correct?
AH: Yes, my co-founder was working as the designer, while I built financial models. He would use algorithmic techniques to adjust the design multiple times a day. I told him, “you realize every time you change this, I have to go rebuild my financial model”.
We eventually integrated the financial model with the design model to provide real-time feedback. As design changes were made, they would automatically be reflected in the financial outcomes. This integration allows architects to gain insights they typically wouldn't have access to.
JF: So by unifying the design and financial performance of the project, you’re holding context in one place for everybody. And giving everyone clarity about the tradeoffs between different design choices. But let me ask the product person question: who does this solve a problem for? Or, who’s the party feeling the pain from the way things work today?
AH: Our primary customers are developers. They want to maximize the value they create on a piece of land. Parafin is designed to streamline the process for developers, eliminating much of the back-and-forth with architects. The current process typically involves:
Developers come across a piece of land suitable for a specific type of building based on market conditions.
They perform back-of-the-envelope calculations to estimate potential returns, without having a concrete design.
If the back-of-the-envelope passes, they contact an architect to create several design options.
The developer then needs to build a financial model for each design variation to make an informed decision about the land's development potential.
JF: How long does that initial loop take?
AH: For very large, complex projects like those I worked on at Forest City, the process could take months. For more typical projects, such as a 100-unit apartment building or hotel, it might take a few days to a couple of weeks to assess feasibility of initial designs.
JF: And how long is that same process using Parafin?
AH: Oh, literally a few minutes. With Parafin, we just need the site location, zoning constraints, and the hotel brand you want to build. Then you can modify your financial model directly within Parafin.
JF: And that’s all possible because of the machine learning (ML) approach y'all took to design generation?
AH: Yes, financial performance metrics serve as the ultimate fitness function real estate developers are trying to optimize for. Because both the design and financial model are hosted within Parafin, we train our algorithms to generate designs optimized for financial performance.
JF: Is there a particular approach y'all are taking or have you experimented with a couple of different methods?
AH: In the process of training the algorithms we use in production, we use what are called genetic algorithms. It starts out with a “gene pool” of designs and generates ~30 designs based upon somewhat randomized parameters. That'll be that first generation. From that group, it’ll remove the worst performing examples and keep the performant ones. Then it’ll begin recombining the performant designs while adding in some random “mutations”. And it’ll go through this iterative process over multiple generations.
JF: And when we're talking about iterating over generations, this is a process happening in seconds or milliseconds, correct?
AH: Right, right. Initially, setting up and training Parafin's algorithms for a new hotel brand or product type can take days. Once trained and loaded onto our servers, computing time for our end user is in the seconds.
Parafin the Business
JF: From a business standpoint, y'all are only doing hotel development right now. What was the thinking behind focusing on that as a vertical?
AH: It was a building typology we could bring to market relatively quickly. There’s a lot of variation with other building types, but with hotels they tend to be very uniform which just made it easier to bring a product all the way to market.
There's also the fact that branded hotels are designed according to brand standards, so there was an opportunity to serve developers who need to build at scale, building 200 properties with some regularity on different parcels across the US.
JF: Makes total sense. In terms of the technology, though, is there any reason y’alls approach couldn’t be applied to all building types?
AH: Hotels are just where we're starting. But there are not too many building types where a product like Parafin couldn't be used. Although currently, highly specialized projects like stadiums, arenas and master-planned megaprojects are more suited for finely-tuned, bespoke generative design systems.
Will Algorithms Eat Architecture?
JF: I have to ask the luddite question. Does Parafin eventually put architects out of a job? Or does it make architects more productive?
AH: Well, we get asked that question all the time. I think it enhances architects by automating the mundane, repetitive stuff. This allows architects to be the architects they trained so many years to be. Racking up billable hours drawing the same buildings over-and-over can suck the soul out of anyone.
JF: And a developer using Parafin clicks some buttons, sees a few viable designs, and then moves ahead with the land purchase?
AH: For me, acting as a developer, I'd still probably have the architect involved very early on. But it’s more of a brief conversation just to sanity check for any red flags. You eventually hand those optimized designs off to the architect to refine and produce the documentation throughout the process. I'd also note that having refined design concepts quickly allows the developer to involve the general contractor in the process much sooner - I always prefer having the GC involved as early as possible.
JF: This sounds like how designers are using tools like Midjourney to quickly brainstorm ideas and generate starting points to riff off of.
AH: Yeah, I think you end up moving forward with something very close - if not identical - to what Parafin generates. But there’s always room for the human sanity check or additional touch an architect might want to add. But yes, it saves the architect a lot of time upfront and they don't have to start from scratch.
Technology x Urban Development
JF: Are there other areas – whether for developers, architects, or whoever – where you feel there’s room for optimization by using ML?
AH: It's all over the place. I could probably rattle off plenty. What comes to my mind first is construction scheduling. You know, handling materials and the logistics of getting the materials to the site, along with coordination of all the trades, each depending on completion of other trades' work. Budgeting and cost control can be highly-complex and begs for ML solutions, especially considering all the money that changes hands from the investors down to the individual tradespeople. For the developers, we're just scratching the surface of what ML can do to help predict development budgets, revenue, and operating costs.
This is an industry built on the knowledge and intuition gained from doing things for years, even centuries, which is invaluable. The quantity of data deeply embedded in the industry is nearly impossible to imagine, and will take decades to be integrated into ML processes.
It's often optimal for the built world to operate off the same tried-and-true playbook, and say "this is how we've done things for my whole career, so this is how we're going to keep doing things." It's seldom optimal to try to build a new way. The build-measure-learn feedback loop in the built world is not like digital technology startups where you can launch something and have users providing feedback all in one day. In real estate development the feedback loop takes years, and the cost of failure is high. That means the opportunities for ML are tremendously deep, and will persist for decades.
JF: It sounds like the upshot here is that by applying ML, we can speed up iteration, because these systems can propose things in rapid succession. And they also won’t necessarily have priors in the same way as a human who’s been doing something for 20 years.
AH: It's a world that relies a lot on intuition and experience. Anytime you can bring data to enhance decisions and speed up iteration, you can arrive at solutions that might not have been thought of otherwise. And progress accelerates accordingly.