Transcript
If you've ever done driving for dollars, you know how labor intensive it is and time consuming it can be, especially when you're driving around neighborhoods in a car jotting down different properties. We've tested this over time -- we actually had some pretty good results with it, but it's just really hard to scale and do it in large volumes.
We later then moved on into virtual driving for dollars with, I would say, mixed results. So we did get some good deals out of it, we did get a lot of opportunities out of it, but it's really hard to also scale. It's people that you need to hire, train, have them go through it, and then at the end of that maybe or maybe not you get good results. It's constantly training, constantly refining, constantly iterating.
One of the questions that came into our mind is, what if we could have AI do this for us? Leveraging these vision models, what if we could actually have it score Google Street View images, give that back to us, and then we could refine down large lists of data to actually identify the ones that appear to be physically distressed? So we decided to build out a pilot of this to see if this would actually work well for our business and something we could actually scale.
I wanted to quickly show you the actual initial proof of concept. So here we're looking at ChatGPT -- you can test this out yourself if you're interested in it, and you can simply just upload a photo. You can see here we gave it a simple prompt: on a scale of 1 to 100, rate the appearance of the house and whether or not it would be a good candidate for renovation, specifically for a fix and flip. One being a complete gut rehab, 50 being moderate repairs needed, 100 being new construction. Pay close attention to the roof, windows, lawn care, etc.
So you can see it gave a very quick response here. It rated a 60 out of 100. I would say that's maybe a little generous, but we did not give it a very comprehensive prompt. And then it gives some commentary on the roof, the windows, the exterior, the lawn. For example, the lawn and landscaping appear to be somewhat neglected -- some work is needed to improve the curb appeal, such as mowing, adding plants, and removing debris. So this was the first proof of concept we tested, and we thought that there was something here.
We then decided to run a broader, larger scale test on this, and I wanted to share some of the results of what we're seeing initially. So let me click over here really quick. This was the first sort of batch that we ran through, and we started to iterate on not only the model that we were using but also the prompting and what we are telling it to give us back in its response.
Here you can see several examples of where we grabbed Google Street View images, we fed it into the model, and asked it to score it and provide a summary. Let's see if I can find a good example here. You see a lot of 50s here -- 50 is kind of middle of the road. Here you can see this one's a 30. So if we look at this, you can see yeah, it's not the greenest grass, the porch looks like it could use some repair work, it looks like the exterior is a little outdated. It gave it a score and says the property appears to be in need of both exterior and possibly interior renovations. The overgrown landscaping, particularly the large tree in front of the yard, suggests a lack of maintenance. The exterior paint looks dated and there are indications the home may have deferred upkeep. The house could benefit from landscaping improvements, exterior painting, and interior updates to enhance its marketability.
Obviously it can't really comment on the interior -- it's essentially interpreting based on the exterior what the interior might look like. But for the purpose of driving for dollars, this is probably good enough.
One of the other things I think that's useful in sharing is actually what the distribution of scores looks like. So in one of our initial samples we pulled a little over 9,600 records and we had the model score it. In this distribution chart you can actually see sort of where the model is landing. Now in our initial instructions we essentially gave it a zero if it can't tell based on the image if it's good or bad -- maybe there's an obstruction, there's trees, maybe there's no house there at all. In those cases we basically just told it to throw it out. Same for minus one.
In these other examples you can see it's sort of bucketing these -- so you see a lot of 35s, 40s, 45s, etc. This then becomes a simple way to take a large data set and narrow it down. We can then start to leverage this to draw the line somewhere. In the example of owner-occupied, which historically for us has been an unproductive list -- meaning there's not a lot of response in the direct mail marketing that we do to that list, in the absence of another list stack like pre-foreclosures or tax delinquency or some other level of motivation -- we're using this data as an indicator of motivation based on the physical appearance of the property.
This allows us to take maybe 10,000 records and draw the line somewhere -- call it 50, 40, or 45 -- where we're actually cutting out the majority of these records that we would then not have to mail. The added benefit here is, as these models get more powerful and cheaper -- and as you'll see with the latest GPT-4o model and the 4o Mini model -- these models are significantly cheaper to actually run. So the actual data cost here that you would spend to refine this list and try to isolate the diamonds in the rough, so to speak, is much cheaper than the mail cost that you spend to just mail everybody here in this list.
So we're still in the very early stages of this. We ran an initial batch of about 3,000 records. We got maybe a little over 30 gross leads off of that. I think we have one under contract so far. We're running our second test currently -- it just went out, it's a little over 9,000 records. And I would say we're optimistic in what this can do for us, and hopefully this can become a core part of our business strategy as we look to grow, scale, and find more areas of separation where you can identify good candidates for property acquisitions.
Hopefully this is another helpful use case of how AI can be applied to your business, especially when it comes to real estate and trying to identify distressed properties. If you have any questions feel free to reach out -- I'd be curious to hear how you're leveraging AI, and if you're interested in exploring this particular use case any further.