This episode provides a technical deep-dive into how LandPricer’s AI models identify and differentiate residential improvements on land parcels. The system uses sophisticated prompting to distinguish between stick-built homes, mobile homes, and RVs with 95%+ accuracy by analyzing multiple visual cues including footprint size, length-to-width ratios, and structural characteristics from aerial imagery.
Key Takeaways:
- Amateurs Automate for Efficiency, Professionals for Accuracy Multiple LLM models run simultaneously and blend results to achieve 95%+ accuracy across different improvement types—prioritizing reliability over speed.
- False Positives Beat False Negatives The system intentionally risks misidentifying unusual structures as improvements rather than missing them entirely, since overlooking a mobile home creates larger due diligence failures than flagging a questionable shed.
- Context Boundaries Prevent AI Errors Training the AI to identify improvements “within the boundaries of subject property” required extensive edge case handling for tree cover, encroachment, and neighboring structures that could confuse the model.
The system translates human due diligence expertise into automated processes by defining specific visual indicators like length-to-width ratios under 2:1 for stick-built homes versus the more rectangular profiles of mobile structures.
(Podcast transcript below)
Welcome to Get Serious. Today I wanted to do another feature update on LandPricer. I know I’ve chatted a bit before on how we were identifying streets and, know, again, the word of the day is nuance and reliability. So taking into account how to identify street access or how to identify a flag lot or looking at contour.
So this time I wanted to do some updates regarding improvements. And for this one, we’ll focus more on the, residential improvements to take a look at. So a few of the different questions, again, this is all going to be on the back end. So the user, you know, for land price, there’s not going to have to answer any of this directly. This is what our AI LLM models, large language models are handling.
through very sophisticated prompting and image analysis. And so we’re trying to determine, you know, for example, are there any stick built residential structures within the boundaries of the subject property? And then we ask a variation that question also related to mobile or manufactured homes, and then also for RVs or any travel trailers, for instance.
You know critical in the way that we’re wording that question as you can notice is that you know Is it within the boundaries of subject property so that that’s even something that has to be? Kind of rigorously defined for the sake of the AI because you know depending on the resolution or the color or you know how close the Improvement may be to another
neighboring property, or maybe there’s a little bit of encroachment, it goes over the boundary line. Like there’s a lot of potential edge cases that could throw off the AI, or maybe there’s some tree cover that can make it a little bit tricky to determine where the residential improvement is. So we have to define all of those particular issues. And again, across all these different image analyses that we’re doing, we’re averaging 95 %
plus percent across blended number of tests. So we’ll run like multiple different LLMs to calculate the actual answer and then we’ll blend it together with kind of best performing. It gets really technical, but that’s the way we’ve designed it to have the highest level of accuracy possible. And that’s just another Sharan Trivatsa type of thing is that amateurs try to automate for efficiency, professionals try to automate for accuracy. So we’re really targeting accuracy.
Again, we want to be the most reliable land pricing tool on the market. That’s, that, is our goal here. And so, you know, also determining, okay, what’s a stick built structure versus a mobile or manufactured home or an RV. So each of those are going to have different implications to the underlying value of the land. Right. And we have different weights or, you know, kind of flags.
from a DD perspective to send back to the user depending on what the image analysis finds. So we have to be able to determine, how do we figure out what’s a stick-built structure versus mobile home and so forth. And so we had to throw a whole bunch of test images and really figure out, okay, the stick-built homes and they on average will have a larger.
Footprint from an aerial perspective the mobile homes There could be you know overhangs or porches for instance, maybe it has like a little bit more of a square structure or We even have like, you know ratios within it for example, like, you know stick-billed homes might have length to width ratio of less than two to one for instance
so we, we try to use all of these various types of, clues for the AI to be, more accurate for identifying the, particular features. and then we have, you know, different analyses that can try to prevent, you know, false negatives and false positives. generally I would,
Rather have a false positive than a false negative For pretty much any analyses that we’re trying to do from an image perspective Not always but in these cases for improvements, yeah, I would rather like miss identify I don’t know some weirdly shaped tree as a stick-built structure and then you can just clean that up later versus like missing a structure Entirely like that’s just a bigger
error compared to accidentally identifying one. So that is what we try to aim for, for those edge cases that might be just a little tricky to determine. I mean, it’s just impossible to think of every single possible parcel in the U.S. and the different resolutions and so forth that can get encountered there. So we just have to make some trade-offs where necessary.
and then how to avoid those false positives and differentiating again, mobile homes and RVs versus stick-built homes and so forth, or even like storage sheds or farmhouses, et cetera. There could be a whole bunch of different residential types of structures. So that’s just to give you another idea on how we are starting to…
really delineate and become very, very accurate on identifying not just improvements in general, but very, very specific types of improvements and all of the implications therein. you know, more complexity within the prompts that we’re doing, but I wanted to, again, kind of give you behind the scenes viewpoint on our overall strategy and thought process behind how we design this. Because again, I’m just…
translating over, okay, what are we doing when we are, you as human beings reviewing land deals? So how do I turn this into an automated AI process as well? There’s a number of other improvement related questions. It’s one of our largest sections that we review, but I’ll dive into those in another episode to come here. Excited to deliver this to you all here shortly.
With that in mind serious land capital for any of your funding needs and land daily diligence Facebook group for zero cost review of your land deals and landpricer.ai Which we just talked about here today. So subscribe and share looking forward to talking to you next time. Take care. Bye
