In this week’s AI Update, David Lykken sits down with Pavon Agarwal, founder of Angel AI, to unpack one of the most pressing issues in mortgage technology: accuracy. While generative AI models like Gemini often “guess” at answers, Angel AI’s deterministic approach ensures precision on critical topics like FHA vs. FHFA credit score requirements. Pavon also shares insights into predictive AI and Eric Schmidt’s bold prediction that user interfaces will soon disappear, replaced by natural conversations with your AI assistant. For mortgage leaders, this episode is a clear reminder that the future belongs to those who harness AI tools that are both innovative and reliable.
[David] Listeners, get ready for another AI update with my special guest and dear friend, Pavan Agarwal. Pavan, good to have you back.
[Pavan] Oh, so much fun.
[David] I appreciate you introducing me to Jennifer. You admire her, you always promote her and lift her up so high, and there’s good reason for that. She is amazing and we had her as a guest and she brought out a good example of how Angel AI is trumping Gemini
[Pavan] Yes, yeah and we’re not just picking on Gemini, there’s no like these general purpose solutions? those general purpose architectures can never deliver the accuracy.
[David] Why is that, Pavan? I think we talked about that. She gave a great example of that when she was talking and we literally did a screen sharing. I loved having Jennifer on. Got to have her back. Thank you so much for opening the door for that to happen. But why is it that that is happening? Because Gemini is supposed to be the end-all, be-all kind of thing.
[Pavan] Well, no, it’s not end-all, be-all. It is a great tool, but you’ve got to use the right tool.
[David] You characterized it as a great tool. Okay, maybe I went a little too far. But you talk to some people and they think it’s the end-all, be-all and that is a mistake, and that’s what we’re trying to bring out.
[Pavan] There’s a difference between hype and reality, right and the hype is that it’s the end, all be all, but the reality is that it’s, it’s amazing. You know, all of these systems have propelled humanity forward, and by you know, hugely forward, and we’re doing things that we never thought was possible before well,
[David] Let’s go into this example you sent me right now, because you just showed an example of someone who’s was asking a question can you use something other than a FICO for FHA loans? and talk about that example.
[Pavan] Let me bring it up here. Let me read it exactly. The question is can I use the Vantage score?
[David] Vantage score yeah.
[Pavan] Yeah, the Vantage score. That’s a new Vantage score approved by the FHFA with FHA financing.
[David] Yeah. And Gemini said what?
[Pavan] Gemini said yes, you can.
[David] Angel AI accurately said.
[Pavan] No you can’t, because FHFA is not the same as FHA, yeah. I think, that’s where it tripped up, because FHFA is a regulator for Freddie and Fannie. Okay and they permitted Freddie and Fannie to use the VantageScore, which is super, that’s great, we need more options and we need new capabilities in the industry all the time. I love that, but HUD has not. So you have to use the right tool for the right job. So Angel AI is deterministic right. It’s not generative.
[David] You talked about that deterministic with Jennifer. That is really good. Again explain to our audience that may not have caught that interview. And, by the way, listeners go check out the link to Jennifer’s interview that’s in the show notes on this one. Go to our website and listen, or just the show notes on this page. What is the difference? What is deterministic?
[Pavan] Yeah. So you know, deterministic. It basically everything you put into it, every question you ask or every fact you give it, it tears it apart, tears it apart word by word, and then it ties those words back to larger concepts. So it builds this internal mental map of what it is that you’re asking and it has to precisely tie this FHA means HUD, which means 3.5% down payment loans and so on, and it’s total scorecard, it’s not DULP and so on. So it creates just that one word FHA. It creates this huge map in its head, all these vectors that connect all together and say, okay, this is what it really is, okay and then it does that for every single word and then from there, then, as it makes this vector map, it then says from this vector map, it says what are the concepts, vector map. It then says from this vector map, it says what are the concepts? and so it goes from the words to a vector map, to a concept, and then from the concept then it says okay, this is what I need to do. And then it goes back the other way to get you the answer Okay.
[David] Very good. That’s a good explanation. A little more detailed than someone may want to remember, but it does give us the importance of the distinction versus. Now, that’s a vector map. It gets a determinant answer, now used in contrast to what ChatGPT or Gemini or any of these other large language models LLMs are using.
[Pavan] Right. And so what they do is they look at the pattern of what the input is. It could be a pattern of words, it could be a pattern of pixels, of an image, it could be a pattern of anything, whatever it’s trained on. They look at the pattern and it says what is most likely the next word or the next dot, or whatever the next, whatever I should generate based on this current pattern. So it takes a look at your current pattern of words and goes and finds in its database the closest matching pattern and then from there it’ll guess the next word. That’s a hyper simplification, but that’s basically what it’s doing, whereas when it’s deterministic it has to take a look, it has to go, figure out what the concepts are and then it has to conclude, it has to distill that down to a single absolute answer. It cannot guess at the answer. It has to be able to figure out what is the answer and it has to be able to explain why it figured, why it concluded to that answer.
[David] So, the conclusion and the moral of this part of our interview today is Gemini. It’s not that it’s bad, but it’s not best when making something, a determination like FICO versus on an FHA, using FICO versus VantageScore
[Pavan] You have to precisely conclude to an answer. In this business you can’t guess at an answer If the AI concludes I could say three possible things, and I don’t know which one is right. It’s going to say I don’t know, and that’s where the augmented intelligence comes in, and then it’s going to loop in a human automatically. Yeah, right, and that’s where the augmented intelligence comes in, and then it’s going to loop in a human automatically. It all happens behind the scenes and it happens like in seconds. It’s like it’s amazing how this thing works and a human helps it along and says yeah, it’s. The answer is number one and every time the human helps it along, now it’s learned. So the next time it’s got a higher weight on number one.
[David] Yep which is what angel ai uses. That’s the human factor and that’s why the human factor. A lot of people say that’s not fully AI, it’s human factor. Yes, you want human factor in these things for the very reason we just exampled. Yeah, another thing you just sent me which is really interesting is the Facebook interview by Ex Google CEO, Eric Schmidt. He’s talking about user interfaces going away, talk about that. That was really interesting.
[Pavan] I’ve been talking about that for you know, since 2019, since we first rolled out Angel AI, because it you know. Again you know he’s talking about it now because what they call agentic AI, which is AI agents, is now hitting sort of mainstream thought.
[David] Agentic is agent.
[Pavan] Agent, right? Not angelic, agentic, agentic, yes, agentic. So we had already, our researchers had already developed this concept and we call them cells. That’s why the company is called Celligence cellular intelligence activate. Have you know? Basically AI-activated cells and AI-driven cells that perform different tasks for you as a human would perform, so instead of you doing it, you have a cell doing it, which is basically what AI agent is that it is a little AI program that does something for you instead of you doing it yourself. And we concluded, you know six years ago now, that if you take millions of these and combine them together, each one doing a little tiny thing, a really, really small thing, just like the cells in your body. Each cell does a little tiny thing, but combined together, right, all those cells combine together and creates the beautiful David Lykken that we see here in front of us.
[David] The old man Lykken, whatever he did, yeah, pretty amazing. It’s really, really interesting stuff. And but why are the user interfaces going away?
[Pavan] Okay. Well then, when we saw that, hey, you can have cells doing everything for you, right, whether it’s sending an email or calculating income on a loan or whatever it is, then why do you need a user interface? You just tell, speak to your computer and you say do this.
[David] That’s the point, yeah.
[Pavan] And then you can say do that and do this and let the point yeah, right, and then you. And then you can say do that and do this and let the cell do it for you, right, what’s the what where? Where is the user interface? It’s just does your personal assistant, who does lots of work for you, right, big, high value task? What’s the user interface to your personal assistant? You speak, speak to him or her.
[David] Yes.
[Pavan] Right, yeah. So it’s the same idea. If your computer is now your personal, complete total personal assistant, you speak to it, right, and you just give it commands and it will do that command right and that’s why at Angel AI, we have a concept called errands, right, because you ask the AI to run an errand for you, right, it is your servant, it runs an errand and completes it and you give her another errand. It’s like an errand boy or errand girl, right, yeah.
[David] When you start looking at it in that context, it’s really amazing how these cells, and I had forgotten. That’s why you named your company Celligence
[Pavan]Exactly, Exactly
[David] Which I’m just going to put a side note in there which has been valued in the multiple billions and billions of dollars. It’s congratulations on the valuation, what you’ve accomplished. And a lot of people start looking at how do they solve these things? And it’s I mean, they’re trying, they’re accurate and they’re just using little pieces here, trying to bolt it together. But you’ve been doing this for over 21 years and that’s why you have something that is as powerful as you do.
[Pavan] Yeah, it’s the time and also the engineering right. Our researchers. Like you know, people will think that, okay, well, okay, so you were ahead, you know, six years ago and now the industry is catching up. So too bad. I said no because our researchers are. You haven’t seen what we got in the lab. We’re ahead 10 years from now. There’s stuff we have. We don’t publish it because why would we let all the trade secrets out? There’s so much in the lab and that there’s so much in the lab.
[David] There’s more to come.
[Pavan] Yes, a lot more to come, a lot more to come. One of the things that I’ll talk about now that is about to come out is a predictive conversation.
[David] What do you mean by a predictive conversation? Is it listening and being able to anticipate where you’re going with that right?
[Pavan] It’s very very human, like OK, so if I say, David, I’m thinking about getting in, I think you’re buying a home, but I don’t have a whole lot of down payment, right, then you would automatically go to the next question would be you know, are you a veteran? maybe if he gave some information, maybe he said that, hey, I come from a military family, I’m thinking about going home and most likely then he’s a veteran, right, and so you could. The AI would piece all those little bits of information he’s got from the conversation and said you know you’re probably a veteran. So, number one, thank you for your service. Number two have you considered a VA loan which you don’t need any down payment? okay so that’s predictive, right? that means it will predict what it should ask you based on the information it has. So you know every other, whenever you take a loan application with any other platform, is regimented, you know. You ask this question, you ask this question, you ask this question, okay, and it doesn’t. It isn’t user driven, It is, computer driven. The computer tells you have to follow its box. The computer driven. The computer tells you you have to follow its box. Okay, so we switch the paradigm around where it’s you are in charge. You tell. Tell me, tell me what’s going on in your life. Tell me what you, what you, really want. I always remember, David, what you taught me is that when you were a loan officer, the first question you used to ask every single client is tell me what you hope to achieve from this transaction.
[David] Exactly what’s right. What’s your why? Why are we doing this?
[Pavan] Exactly right, exactly that’s good, you remember that yeah.
[David] It anchored it in. Once you get that anchored in, probability for success goes up exponentially.
[Pavan] Exactly, and that’s the whole idea of the AI. AI is going to start from your why. What are you doing this for? What do you hope to achieve? And then it will ask you the right questions, right?
[David] Which goes back to the deterministic nature of what you’re doing, which gets really exciting.
[Pavan] This is not. This is technically predictive now.
[David] Oh, that’s okay. So predictive versus deterministic. Contrast those for us.
[Pavan] Okay. Deterministic means there could only be one. Okay, it could only be one answer, one conclusion. Two plus two can only be four. Okay, probabilistic means I’m sorry. Yeah, predictive means that that we’re going to be talking about AI today in this podcast. You sent me a text saying, hey, let’s do another podcast, then I assume we’re going to be talking about AI.
[David] Yes, so it reads into it and it gives you right.
[Pavan] That’s predictive. So that means if I was the AI, right, you sent me a text saying, hey, let’s do a podcast today. And if I was the AI, I’m going to predict hey, David probably wants to talk about AI and I would respond with oh, I just saw this great thing from Eric Schmidt, founder of Google, talking about the same things that we were talking about years ago. So, predictively, I’m assuming you want to do a podcast about AI, right, and I respond to you with a suggestion of what we can do in the podcast. That’s very human like
[David] So when you come to underwriting, we want to be very deterministic. When you’re in a relationship of a loan officer, you want it to be very…
[Pavan] predictive, predictive, right.
[David] Predictive, predictive of suggesting it. That’s a really good distinction.
[Pavan] So that way you can help guide. It makes the user more comfortable. You can guide him towards a good answer. Right, if you keep them.
[David] So how do you then shift between the two, predictive versus deterministic, when you’re in the transaction, when the Angel AI is actually working? What part is working and how do you begin to make that happen?
[Pavan] It’s really simple If the user is asking for information.
[David] Simple. I’m jokingly thinking simple for you, yeah, but not for us.
[Pavan] Well, I’ll put it this way the explanation is simple, right? If you ask the computer for an explanation or for information, it better give you exactly right. That’s deterministic, right now, when the computer asks you for information, that’s predictive. The computer has to ask you a meaningful question an applicable question.
[David] So okay, it’s an important distinction. I want to get into that more, but I’m trying to keep these AI updates down to a shorter segment. So you recently spoke at NAREP. I was really excited about it. I read over your presentation. It looked like you crushed it there. Why did you get asked to speak at NAREP and what was the takeaway that you walked away from that?
[Pavan] Yeah, so that relationship started at the US Strategy Conference we did in DC in July.
[David] Yeah, Gary was there.
[Pavan] Yeah. So Gary was there and I interviewed him and we started talking and he met a lot of my friends and business partners. He met a lot of people in the key people, top people, leaders in crypto and AI space at that end and he’s been thinking for a while, how do I expand NAREP into crypto and AI? And he had been reaching out and looking for speakers and he didn’t have any. He wasn’t having any luck and then after we met and he saw the talent that was in that room so he asked us to put that together. So we were invited by him to put together the crypto AI track for NAREP and we pulled in some of the greatest minds across the world for the NAREP. And they’re nice people, I know them well and they did it as a favor, like, for example, Lauren who runs an organization called UTXO, right, and they.
[David] What’s that stand for?
[Pavan] Actually, I’m not sure what it stands for, but they’re the largest Bitcoin association and, and you know, their conferences have 50,000 people at a time and so he’s used to speaking to 50,000 people, and he agreed to do this small conference with a few thousand people, you know, and a hundred people.
[David] You were able to get him to come in and speak Wow.
[Pavan] Yes, and it’s because of the relationship. You know, business is always about relationship.
[David] Yep.
[Pavan] Right. So these are the level of people we brought to the conference. And then, of course, Gary has made his own twin and he’s got his twin up and running in Angel AI and we also created the NAREP Angel AI version where, if you go into askangelai/avance, which is what their app, their NAREP app, is called Avance A-V-A-N-C-E Right, okay, so if you go in there and you’ll see, Gary’s twin will come and talk to you and you know and you can start doing various things with it.
[David] Oh, that’s, amazing.
[Pavan] Okay. So that’s kind of how it all started. And then so I so these wonderful speakers came. And so these wonderful speakers came. They were so generous with the time they came and they spoke. We also had Kyle, and Kyle Wiley he’s amazing. He was in Trump administration, number one senior guy at the Department of Energy organization, and he’s probably going to be appointed to the FTC on the crypto, head of the crypto committee and Trump 2.0, and so he was just an incredible wealth of knowledge. I mean is like that, the you can, like he could, you can. When you have a conversation with him, you could peel the onion as deep as you want and he can answer everything and he’s open. He’s open-hearted and doesn’t hold back information. Whatever you ask him, he’ll tell you.
[David] That’s one thing I’m finding w ith the crowd that you run with, you draw people in or you attract people around you that are very much about giving. We just did a podcast with Donn Healy. I love that story and how you are partnering with him, and that’s a whole other side of the industry. It’s really working almost at the microfinance. It’s almost working at the artisans, bringing the people that are doing the small little tasks off in some obscure country and bringing them to market through what you’re doing with Havona. That’s very exciting.
[Pavan] Yeah. So what he said was you don’t need a user interface anymore. Because his literal words were you can say you know. Hey, AI make me a user interface with whatever buttons I need to solve this problem. Okay, and the AI would generate that user interface. But I think he missed the mark slightly on that. You don’t tell the AI to generate a user interface so then you can use the user interface. You just tell the AI. That’s kind of silly. You just tell the AI just do it, just bleep and do it for me. You don’t need to tell the AI to make the user interface so that then you can press a button on the user interface to do the task. That’s kind of silly. The whole concept, the whole principle of Angel AI is you just tell her just do it Whatever it is that you want done and she’s just going to do it. There’s no need Tell her, just do it, whatever it is that you want done and she’s just going to do it, right, so there’s no need for buttons or user interfaces anymore. You just explain what you want and say, and then.
[David] Simple, plain English.
[Pavan] Right, simple, plain English. You’ll go and run and do it Just like you would tell your assistant. Right If your personal assistant you would tell you would tell that person the same thing.
[David] Yeah, it’s fairly fascinating. I think the old saying a lot of people still have the fear that AI is going to take, eliminate their job. No, what’s going to eliminate your job is someone using AI or using it better than you do. That’s why we’re doing these AI updates. Pavan, so good to have you on. I look forward to doing another one with you next week and kudos to you on the success that you’re having. How many users are you adding a month right now? I mean, you’ve been doing this month over month.
[Pavan] Yeah, I mean 15,000 plus users a month. We’ve crossed 250,000 now and it just keeps on growing. And again, no media.
[David] That’s without any media, it’s just word of mouth,
[Pavan] Just word of mouth, and attending some conferences and some e-blasting. Our marketing budget is like nothing.
[David] I’m telling my clients you better get up on Angel AI. If you don’t choose to use it, that’s your own business, but your loan officer is bad. But here’s what would be a real embarrassment, when a realtor is using Angel AI is getting more accurate answers to underwriting questions than your loan officers giving or your underwriters giving. That’s going to be a problem, that’s going to be a big problem for you, so get used to it. More and more and more people are signing up for it and they’re going. This is amazing.
[Pavan] Yeah, and what I found at NAREP was, because there’s a lot of big, a lot of brokerages were there, big and small, right.
[David] Yeah.
[Pavan] And you know like some of them are considering building their own AI, until they saw what Angel did Like you already do everything, okay. And so they’re like, well, should I build my own AI? And I said well, yeah, you know you can okay, it’s gonna you’re gonna spend a lot of money, it’s gonna take a long time to catch up, but the other problem is that your agents won’t appreciate it. Okay, the reason they won’t appreciate it is because they have. They have an option like Angel AI which is theirs. Is there a personal angel as your personal AI? Okay, so that means, if you, if you move, if you’re a realtor and you move from, say, LPT to Epic or to Century 21 to Remax, right, your angel AI, your personal AI, stays with you everywhere you go and all the content it generates with you stays yours. All the listing sites, all the webpages, everything that you’ve trained the AI on to do for you as your agent, stays yours, and all of your training that you give Angel AI is you own and it’s in your wallet, because we create a vault for you and your personal training goes in that vault and only you have the keys to it because it’s your blockchain wallet keys. Okay, so I’m like you know. So the problem is, you’re going to be you. If you’re trying to build your own AI, you have to compete against. You have to convince your agents to use yours versus something that will be theirs.
[David] Very good. We could go on and on. I want to pick up where we just the where you were just going on our next AI update. So, listeners, be prepared for next week’s AI Update. We’re going to pick up right there. We’re just going to continue bring you knowledge each and every week. Pavan, thank you so much for being here.
[Pavan] Thank you, David
[David] Awesome. Folks, this episode was sponsored by Angel AI and I’m grateful for them and all they’re doing, the innovation. Check it out. Go to angelaicom or actually start using it by going to askangelai. Again, Pavan, thank you, appreciate you being here.
Important Links
- Podcast interview with Jennifer Vallinayagam
- SunWest Mortgage website
- Angel AI website
- LinkedIn – Pavan Agarwal
Pavan Agarwal is a renowned leader in the mortgage lending industry and a pioneer in bringing artificial intelligence to the financial markets. Agarwal serves as the President and CEO of Sun West Mortgage Company and Celligence International, LLC.