by G. Sax, Head of Communications, RESO
Welcome to “Three Questions,” an interview series that introduces you to real estate industry professionals, their businesses and how they interact with real estate standards. The goal of the series is to humanize the tech side of the industry, fun included.
This week, we chatted with Nathan Brannen, Chief Product Officer at Restb.ai, to talk about – what else? – artificial intelligence (AI)! But not just AI. We also delved into European real estate, camera quality and evolving a product through continuous improvement. Enjoy!
Q1: Everyone is talking about AI in and out of the real estate industry. This must be exciting for Restb.ai, which seems poised to help people navigate this ever-growing technology. What are the primary challenges in instituting this type of tech in a venerable vertical like real estate?
Nathan: It’s been quite interesting for us to look at the evolution of AI in proptech in the context of before and after ChatGPT. Before its release, people weren’t doing anything with AI in the real estate space, and now, suddenly, everyone is!
What we’ve learned from our past ten years of working in AI is that it’s easy to sell the dream of what is possible with AI, but actually getting it to work is what is challenging. There are always unexpected things you have to figure out to make it actually work as intended.
A simple example for us is the automatic population of details in real estate listings. Let’s imagine our model detects concrete floors in a photo. Does that mean the property has concrete floors? Well, it depends on whether the photo was of a garage, basement, living room, etc.
Will the appliances detected in the property’s photos be sold with the listing or will they go with the owner?
When we signed our first Canadian client, we had to retrain all of our models to account for snow, as the images were unlike anything our models had ever seen before!
It’s often impossible to predict what corner cases you’ll run into and have to account for prior to creating a solution. Fortunately, our clients have been understanding that the tech evolves, and it is always improving.
In fact, our clients are among our best assets to identify and resolve gaps in our models. Finding new ways to improve isn’t the issue.
As part of our continuous improvement processes, we do about three to five model iterations per week, all to adjust for the little things that we find. We’ve dedicated enormous resources to build out a tech pipeline to be able to quickly iterate and resolve the issues and opportunities that we come across.
The RESO Data Dictionary has actually been a great reference point for us. It helps provide a base definition for what we’re defining in our models, even if there are still some topics that may be more open for discussion than we initially realized.
For example, a bungalow in Canada is a ranch in the U.S., but a bungalow is a different thing than a ranch in the U.S.
A porch in the UK is different from a porch in the U.S. A porch is kind of like a mudroom in the UK.
In Brazil, they have these rooms that just have a sink, which our model really had a tough time with. So we worked with a couple of Brazilian portals to help us resolve that.
Q2: You travel pretty far to attend RESO conferences, coming all the way from Barcelona, Spain. What is it about this particular conference that attracts you the most, and what can RESO do to attract more attention from other European entrants into the real estate standards space?
Nathan: What attracts us is that RESO helps us understand real estate at the data level. We’re an AI company, so it was nice to find people speaking our technical language.
Most of the conferences we attend are more sales oriented, which means we’re having to sell the vision or dream of AI. However, with RESO, it’s much easier to talk to people about how it all will actually work. People don’t have their guard up, and it has really helped us refine our approach with MLSs.
Additionally, the size of RESO conferences is approachable. It’s the only conference where I know the employees of the organization putting on the show. It is the type of organization where people really seem like they are there to help you, and that is refreshing.
To answer the second half of that question, you have to understand that European markets are so different. Standards are a big issue, for sure, but an even bigger issue is that one company can control an entire country’s market. Most other countries, such as Spain, feature pay-to-list models. That is, an agent has to pay to list their property on the portal. If they want to list their property on multiple portals, they have to pay multiple times. This inevitably leads to one portal having the largest amount of listings and, therefore, little incentive to drive standards that serve the entire real estate community.
We work in many markets outside the U.S., and that has instilled an even greater appreciation for the collaborative approach we see at RESO and within the MLS world. While many people at RESO conferences may have complaints about various things with U.S. real estate, I can assure you that the alternatives we see in other markets are much worse.
Q3: How much of a factor is camera quality in determining things like property quality and renovation status through AI, and has the advancement of camera quality on phones significantly changed your product offering over the years?
Nathan: I think there is a misconception that overall camera quality has gotten much better, at least across the board. I’m continually shocked by the low-quality photos we receive from many markets, including in the U.S.
Then, when we started working with appraisal and inspection companies, we realized that these photos are even worse than MLS photos. It quickly became critical for us to ensure that our models could ingest and handle low-quality images.
We built a model that determines how bad an image is, accounting for blurriness, darkness, overexposure and other factors. This allows us to audit our other models to make sure they work well on all types of images.
Particularly, when we’re doing things at the property level, such as scoring a property’s condition or quality, we will weigh the quality of each image to determine how to account for it in the overall score. We may even potentially throw it out.
If a property has 30 images and two look like they were taken while running in circles, we can minimize the weight of those “low-quality” photos in the overall assessment of the property.
I like to say that when we started, we were basically a photo company that analyzed photos.
But we learned that people don’t care about photos as much as they care about what photos can tell us about a property. We needed to become a company that could provide property insights.
Imagine a property with an unfinished basement. No one would judge the condition of that property based on the “condition” of the unfinished basement. They would be more interested in the condition of the finished upper floors.
At the end of the day, we’re trying to train the AI to analyze a property in an unbiased way that provides the most usable and structured data without our clients having to worry about all the “what ifs” and corner cases.
