Vin is the founder of Foxy AI, a prop-tech startup that uses artificial intelligence to look at photos to determine how much a property is worth. In this episode, we have a high-level discussion of how the software works and how it will revolutionize the way we use software to evaluate properties at scale.
Anything is possible as long as you have the right team to execute. Even though Vin had no experience with artificial intelligence, he had a great concept and was able to hire out more specialized tasks. Foxy AI is an incredibly interesting concept and I’m excited to see where it goes in the future.
2:01 – Vin got started in real estate pretty much by accident back when he was in college when summer internship in Boston fell through at the last minute. He had already rented an apartment for the summer and he wasn’t sure what he was going to do. And a great friend of his had just opened up his own real estate brokerage, and he offered Vin a job as a rental agent. That sounded better than flipping burgers and real estate is certainly a more useful skill so he got his license and started renting apartments for him and he hated it. Vin didn’t like peddling these crappy one-bedroom apartments in Boston and to be honest, he wasn’t a great salesperson but he really liked the concept of real estate. So he suggested they start a property management company to go along with the brokerage. And the idea was to sell an investment property to high net worth individuals with the pitch of they have the rental brokerage and a management company so you don’t have to do anything except cash your check. And this went really well for them.
7:57 – Vin found his data scientists on Upwork and found them in Canada. They were great for the initial process of collecting and cleaning and organizing that data. And then when he met his partner, Frankie, one of the reasons that he was so interested was that Vin already had all this data together. He thinks a lot of people that are in AI don’t really want to deal with the data collection process because it’s time-consuming, it’s tedious, and it’s boring. So the fact that Vin already had all of this data together, definitely helped with getting Frankie on board. And that’s why Frankie said, “Hey, do you mind if I take a look at this data and play around with it for you know, a month or two, see what I can do”. And that’s what he did. So Vin would say, from the time that he started with the data scientist to the time that they’ve put out our first model was probably about six or seven months.
10:59 – Vin’s team consists of 3 members. Vin, his head of AI and a junior engineer to support him. They outsource a lot of the smaller tasks that need to be done and as they grow up, they will bring on more people to the team.
11:39 – The way that Vin train the neural networks is to look at photos and they produce what they call an image feature vector. There are kind of two different approaches. One approach would be to try to identify features within the photo that are correlated with value so you might try to identify stainless steel appliances, or granite countertops, or hardwood floors. And when you begin to make a list of all of those, not only positive indicators of value but then also all the negative indicators of value. So things like scratches or holes, water stains, this list becomes very, very long, very quickly. Instead, what they do is we take a global approach to the photo. And this is where the power of deep neural networks comes into play, where they allow the neural network to identify features within the photo that are correlated with value. And the neural network then produces what’s called an image feature vector. And the feature vector is is a series of numbers and the best way to describe it or understand it is that when they’re dealing with valuation models, it’s a mathematical model. And within that model, you’re going to have many different types of features.
17:55 – Vin charges people on a per API call basis. Their goal is very simple. It’s to be a back end provider of visual property intelligence APIs that companies of all shapes and sizes can build into their workflows, their websites, and applications to extract data from their property photos.
19:04 – Single-family investors are using the condition scores in a few ways. One is if they use automated valuation models, they are incorporating the condition as another feature in their model to improve accuracy but not everyone uses ABMs. So another way they’re using the scores is to filter investment opportunities.
22:31 – Vin is trying to build a whole new category of what they call visual property intelligence in the future. So extracting information from images for use in applications like condition scoring and valuation models.
23:15 – Some of the challenges that they face are challenges that really everybody in artificial intelligence or machine learning really face, and that’s acquiring the necessary data to build out these models. Data in real estate is still very siloed, it’s still very protected. You can’t just google property data set for real estate or something like that and get enough data to build really robust models. So it takes a lot of time to scrape lots of data and combine that all.
24:37 – Vin explains Data Augmentation. It’s when you take an image, then you can flip it and it’s like you basically doubled your data set.
27:08 – Vin met his business partner through his friend, Jeff, an entrepreneur. They were talking about this idea and Jeff said, “Hey, a friend of mine that had done some work for one of my startups. He is an AI researcher for Crimson hexagon and I definitely think that you should talk to him, see what he has to say”. That’s where Vin met him. He showed him the data that he’d already had collected and he was super interested, he was like, “let me take a look at what you got here and see what I can do”.
29:21 – The way to test the accuracy of Vin’s vector box is to put it into your evaluation model and compare it to the valuation models the previous accuracy. They did some of their own testing, where this was before Zillow officially rolled out their product. But what they did was they scraped a bunch of properties that were for sale on Zillow. They got their listing price and they got all the photos. They then built a new model to combine their data with Zillow’s estimate, then they predicted a new valuation for the property of its new sales price. They waited for those properties to sell and they then compared their prediction with Zillow’s estimate with the sale price. They broke the properties up into three brackets and properties that were less than $175,000, their prediction outperformed Zillow’s predictions 60% of the time. And for properties that were over $175,000, their prediction outperformed Zillow’s estimate 40% of the time. So just by adding in this image data into a very, very basic model, where they were combining their image data with this estimate, they were able to produce a prediction that was closer to the final sales price more often than Zillow’s estimate was.
32:24 – If Vin raised money, they would probably build out more of a team. Bring on a couple more engineers to kind of speed up the product development pipeline, and also probably bring on a data engineer to help them manage not only data sets that they are collecting externally but also data sets that they are putting together from kind of incoming data streams.
Clint Coons is one of the founders of Anderson Business Advisors, a firm that specializes in creating asset protection entities…
Justin is a real estate investor who has done almost 2000 deals across the nation and in this episode, he’ll…
David is a real estate investor and a real estate coach. He has been investing in properties for almost 20…
Andrew, a real estate investment developer, is the owner of IronGall Investments, an Austin, Texas-based real estate development company. They…
Chris is the President and CEO of Smart Growth Inc., a California-based real estate and development firm. They are focused…
Rafael is a real estate coach and an organizational psychologist based in Miracle Valley, Arizona. He owns several real estate…