Revealio: AI-Powered Property Condition Scoring | Production-ready AI Agents Hackathon @ AIBy theBay
So first off, thank you. We're very excited to be here. We have started with this idea at Ravilio. Ravilio came from my background. I am actually a real estate investor who wants to be a techie. So with that in mind, one of the things as a real estate investor that you do is you go around, drive around the town and you look for distressed properties. Well, as you can think, towns can be pretty big. That takes a lot of time and a lot of resources
What about if you want to then find properties in another state? So, what we've done is we have taken AI and automated this. And so, what that looks like is as follows. One is we connected with with the Google API, Google Maps API. What we had to do is we made it so that we could take the uh street view pictures and then put them into our agent our multi- aent agentic system and then grade those pictures and create a distress score for that property. So how we did that is so we we built it out we brought in the the street level view and then we ran in two different LLMs. So one is anthropic and one is open AI and both would grade the property picture based on a weighted scale because certain things are more important to me as an investor than other things. Like if there's a a a tarp on the roof that means okay that's a pretty distressed property. So we did two agent or two LLMs as well because what about for if the scores didn't match up then we'd have it go back and they'd rescore
So that's a high level of how it works. And then from there, so we got all that working and then you as a as an investor then go into our site and actually put in what address or what city and then buy a list from that. Um, so I'm going to turn it over to my partner Jenna. Sorry it's uh not being stable, so we'll do our best with that. >> It's going to be pretty pretty challenging. I can't really show anything. Um well as Bryce mentioned he talked through the business problem that we were trying to solve. So in an effort to solve it we built a multi- agent agentic system that solves this problem
>> Um so this is an overall architectural diagram of the agentic system. You can see that I'm pointing you guys. Um over on the left hand side, that's where the user sits. They're interacting with um the what's that? Oh, so we have an orchestrating agent that the user interacts with primarily. They're going to input something like an address or a zip code in version two zip codes. Um and that that orchestrating agent decides what to do with that information. So it'll call different tools. Got to breathe
Um, thanks. Calls different tools based on what it needs to accomplish. So, if a user puts in an address and um hits the assess property button, we know that we're assessing properties for a grade. So it will call the vision agent which will reach out to the Google maps API and pull the street view image and it takes that image and sends it to the grading agent which analyzes that image for based on the rubric that we created for the grading. It then sends it to that's the scoring agent and then the explanation agent is what actually does the um I have to read my own stuff. Explanation agent is what does um the call out to the second LLM. So it calls out to anthropic to see if we can get a second opinion on the grade that it got. If the grades are very different then we need to resolve that delta
We'll actually just send it to both again and see if we can't get a similar result right now. But the idea is that we're validating the grading so we know that it's accurate. Um, we have incorporated a Neo4j vector database to stash those property grades in and at a later time we'll be able to look at that vector database and call and see uh relations to properties. So if somebody wanted to see properties in a certain neighborhood, you know, maybe decide where they were going to drive that day and look for disheveled properties, they could focus on um this is actually kind of challenging. They could focus in a certain area. So here you can see the Neo4j u vector database in its little infancy starting to form relations for these properties that we're testing. This is actually the um AKA uh orchestrating agent health API because I wasn't really able to get um a UI in front of AKA. And let me pull up this guy
So I'll run the demo now. I just want to pull up Come on, you can do it. Yeah, I'm just going to maximize um I have a T-Ux session going uh looking at the Red Panda streams. Um so we're looking at four different streams. We're sending the events, inter agent communication failures, the the API calls, all of that to red panda, not that. So, it's going to happen pretty quickly once I run this um property, but you will be able to see the counters on the stream events uh change. No, >> I'm going to use a known good address is what I'm going to do. >> Yeah, the headquarters for Google
That wasn't very sexy, but hopefully some of you guys saw the stream events hitting Red Panda and you can tell things are happening or we're monitoring. Um, these are the property details that we got back and you can export your list to CSV. So that's it. >> Yeah. >> And then just lastly, we had one other part was an API with uh with a one of the largest property uh detail um holding companies in the country. basically Fanny and Freddy, they use their data for doing refinances as mortgages. We've actually we had a they had an API issue on their end this morning, so we weren't able to connect it in. But once uh that's cleared up, what we're able to pull in is basically property owner, their address, the the loan amount, um if there's like a pre-foreclosure or any of those kinds of things, that all come out on this Excel on the CSV as a deliverable for the people that buy these lists
So, that's going to be our next step. again uh we implemented all this stuff for the first time aa and all these and so we've appreciated that but that's it. Thank you.