Episode 5 of The Applied AI Podcast

About the episode

Jason DuPont, COO of NEXA Mortgage, discusses how AI is changing the mortgage industry through practical implementations and strategic partnerships. The conversation reveals how one company processes $10 billion in loans annually while pursuing aggressive AI adoption across the entire mortgage lifecycle.

The future of loan officers as influencers

DuPont predicts loan officers will evolve into influencers who drive traffic rather than process paperwork. Drawing parallels to his 1996 start with manila folders and carbon copies, he sees AI assistants handling calculations, documentation, and technical work while humans focus on relationships and trust building.

This shift mirrors broader knowledge work trends. Family attorneys, doctors, and other professionals will similarly transition from technical execution to counseling and influence roles. The technical knowledge requirements diminish as AI handles complexity behind the scenes.

Nexa's Tidalwave partnership

Nexa partnered with Tidalwave as their first client in the third party origination space. The platform starts at the point of sale (1003 application) and uses agentic AI to guide customers through the mortgage process.

The system ingests partial information from any source (a loan officer taking notes at a restaurant, incomplete phone conversations, scattered documents) and intelligently completes the application through natural dialogue with customers. It determines optimal loan products (VA, FHA, conventional, non-QM) based on customer profiles and lender requirements.

Tidalwave's architecture accounts for lender overlays and underwriter variations. When one lender has specific restrictions another lacks, the system routes accordingly. It tracks patterns where seven of eight underwriters approve certain income calculations, building case precedents for future applications.

Personal AI experiments reveal enterprise potential

DuPont built personal health tracking systems using n8n orchestration, combining genetic data, quarterly blood tests, DEXA scans, and 1,534 days of Whoop fitness data. The system identifies contraindications between medications, optimizes workout recovery, and personalizes nutrition based on genetic markers (including a missing gene that prevents him from tolerating smoke exposure).

His recruiting system demonstrates enterprise applications. Outbound voice AI contacts loan officers and routes them to Thursday Zoom calls about mortgages. These AI calls are recorded, analyzed by another agent, and the prompting continuously improves based on performance. Virtual human assistants then engage qualified candidates before DuPont personally connects.

The crypto trading agents he developed go further: a CEO agent creates subordinate agents for specific tasks, with an auditor agent validating outputs before execution. This hierarchical, self-improving architecture shows how enterprises can deploy AI that evolves without constant human intervention.

OneTouch close: the 54-minute milestone

Nexa achieved a 54-minute window from loan submission to clear-to-close approval—not conditional approval, but full clearance. This benchmark, accomplished without AI, sets the target for automated systems.

Every loan now undergoes analysis to identify friction points. Why does this condition exist? How can we eliminate it? Each partnership and technology implementation targets specific bottlenecks in pursuit of scaled OneTouch closes.

The technology stack beyond large language models

OCR technology handles document ingestion. Classification systems categorize documents. Indexing enables rapid retrieval. Pattern recognition identifies fraud risks. Machine learning algorithms detect anomalies. Each technology serves specific functions within the broader ecosystem.

DuPont emphasizes that true enterprise AI requires orchestration platforms, not just language models. Companies claiming AI capabilities while secretly using human backup will lose to those building genuine automated systems. The mortgage industry's complexity—regulations, overlapping processes, redundant workflows—explains why other sectors adopted AI faster, but also why the opportunity remains massive.

Practical implementation strategy

Organizations should identify where current approaches genuinely fail. Look for problems involving multiple data types, requiring transparent decision-making, or demanding capabilities that evolve faster than vendor roadmaps allow.

Start with pilot projects proving value before scaling. For mortgage companies, this might mean automating document classification before tackling underwriting. Build toward comprehensive intelligence only where complexity delivers returns. The OneTouch close represents the ultimate goal, but incremental improvements in processing time create immediate value.

Cognitive hive AI architecture and Talbot West's approach

The discussion validates Talbot West's Cognitive Hive AI (CHAI) framework—modular, interoperable systems where specialized agents collaborate. DuPont's implementations mirror this approach: agents that audit other agents, systems that improve themselves, architectures that remain flexible as capabilities advance.

Rather than monolithic platforms creating vendor lock-in, the future belongs to orchestrated intelligence. Companies need architectures that swap components as better options emerge, combine multiple AI types for complex challenges, and maintain transparency for regulated industries.

The mortgage industry stands at an inflection point. Those who surf the wave will capture disproportionate market share. Those who resist will feed the growth of early adopters. The technology exists today. The question becomes execution speed and strategic courage.

Episode transcript

Jacob Andra: Welcome to episode five of the Applied AI Podcast. I'm your host, Jacob Andra. I'm also the CEO of Talbot West, an AI enablement company. Today's guest, Jason duPont, is the COO of Nexa Mortgage. And is a wizard at AI, both within the mortgage industry and some interesting applications he is using AI for in his personal life as well, which he'll cover in this episode.

There's plenty here for mortgage professionals to be interested as well. But additionally, anyone from any industry can adapt some of Jason's insights and the concepts we cover to their industry. So there should be something here for everyone. Enjoy.

I'm here with Jason duPont of Nexa Mortgage and Nexa is using AI in its different operations in some very interesting ways. Jason, why don't you introduce yourself, talk a little bit about yourself and then we'll dive into the AI side of things.

Jason duPont: I appreciate you having me. Be here. Goodness. Been in this industry since 96, so a while. I've been with Nexa about a little over five years. Five and a quarter, something like that. Right. It's fun to be here. It's fun to grow. It's fun to see others grow.

Jacob Andra: Yeah, give us a sense of Nexa, its size, its growth rate, you know, anything else interesting about the company?

Jason duPont: You know, right now Nexa is on pace to do about 10 billion. This year we have about, we, I think we just hit 3,300 loan officers, and our goal is to get to 15 billion. So our decisions that we make, what we're going after is to hit that 15 billion mark, and that's our next threshold. Doesn't mean that's our destination. That's just along the journey. Right? We're not destination based. We're journey based.

Jacob Andra: I like that. I can relate to that, and that's a nice ambitious goal. But like you say, onward from there, once you hit it right.

Jason duPont: Yeah, for sure. For sure.

Jacob Andra: Cool. Well, let's talk about AI in the mortgage industry. I know that Nexa has partnered with a couple of different AI companies, and also in yours and my conversation in the past, we've discussed some interesting ideas about AI and how it applies to the mortgage industry. So maybe we can just cover these partnerships, what they're doing for Nexa.

Also, I'd just like to hit some of the main sorts of areas where AI can make a difference across the mortgage life cycle. There are different swim lanes and different sub, I guess you could call them sub-disciplines or departments within the mortgage industry where different types of AI apply in different sorts of ways. So maybe we can just kind of dive right into some of that.

Jason duPont: Yeah. I mean, where do you guys see, where do you see AI being three years from now in the mortgage space? Right. We're talking in the mortgage space.

Jacob Andra: Yeah, well certainly, anything involving document processing, document review or document generation is gonna be gobbled up by large language models. Copywriting is going to probably be a combination of different AI technologies collaborating together to make that, sorry, underwriting. I think I said copywriting there. Underwriting's...

Jason duPont: Copywriting either. It's gonna be hard to distinguish right between.

Jacob Andra: Underwriting, underwriting is gonna be served by some predictive models, machine learning, and probably some large language model wrappers, kind of collaborating together and so on and so forth. I mean, I wanna hear your opinion because you're more in the mortgage industry than I am, obviously. What do you see? What are these partnerships and what's going on with AI in the mortgage space?

Jason duPont: Well, I think, and the reason I wanted to start out just kind of reflecting the question a little bit is for me. And this is a big thing that leadership at Nexa believes, not just myself. And I think a lot of people in the space believe this, but we'll find out. I believe loan officers are gonna become more just like influencers and they're gonna be the ones driving the traffic.

So I liken it to when I first was doing loans back in 96, I mean, we had these manila folders, we had these little clips called echo clips. Everything was carbon copied. We'd hand calculate our good faith estimate, or we can calculate, you know, or write out the good faith estimate. I mean, or hand calculate APR. I wonder if loan officers will even know how to calculate debt ratios in the future. Why would they? Because they'll have their assistants, their AI assistants do it for them. And I think that what you're gonna do is you're gonna see some people get licensed or whatever, and you'll see an influencer on TikTok that's just driving traffic to their assistants and helping them do loans. That's high level. Go. Yeah, go ahead Jacob.

Jacob Andra: No, I concur with that. I very much agree with that prediction, that future. And I see that a similar thing unfolding across different knowledge work areas where humans will be more the points of influence. And also for people who want that human connection. I mean, imagine like doctors, attorneys, imagine like family practice attorneys that, you know, they don't really need to know the law to the same degree they did in the past. They don't need to spend a lot of time doing the work that they're currently doing because AI technologies are doing all that for them. But they are acting as sort of counselors, therapists, influencers, et cetera. So playing that sort of human face, human role. And I do see what you're saying that loan officers will also be doing that.

Jason duPont: Yeah, for sure. And I think it's gonna be all facets of the industry are gonna change very rapidly. Look, I don't think that loan officers necessarily need to worry about their jobs. They need, they do need to embrace, they use the analogy of like, the waves coming, right? This big tidal wave, what are you gonna do? Are you gonna let it crash on you? Or are you gonna surf? Well, I want to surf it. I think that's gonna be fun, and I hope others are like-minded in that. I know you are, because that's what gets you to that next level.

You asked about partnerships, so not deflecting on that. I just wanted to kind of get some foundation there first. So the biggest partnership that we announced was through Nexa and Tidal Wave. So Tidal Wave is the first one out there that we've seen, and we've met with a lot of companies that's really putting into that buzzword of agentic AI. And they're starting from the front end. So a lot of these companies are starting from the back. They're starting from the funding or the closing. They're starting from the front, from the point of sale. So they've started a point of sale system and Nexa is their first client in the TPO or broker space, the non-delegated correspondent, all that space. We work with them every day. We have 30 beta testers on it, all licensed loan officers, and we're just growing with them. And then where it's going from there is, it will go front to back and it will be where the loan for the consumer, the process for the consumer will be a lot easier and the loan officers will slowly become, like we just mentioned, influencers.

And we really see that. We really see that piece of it. There's a lot of technology out there. And just to segue real quick, you mentioned different industries. So like, my health, I've done genetic tests and I do quarterly blood tests, annual tests, scans, all these different things. Right. They're all in an AI system that I made. And so I can put in there and say, Hey, look, this came up on my test and it can tell me, well, you're taking this, this conflicts with this. So contraindications, all those things are there on a very personalized level. And even like workouts, like it can say, Hey, my body type doesn't do well on maybe strict keto because of my genetics, or here's an interesting one. I have a missing gene that makes it so that I can't be around thick smoke, so I can't eat supercharged food, or I can't go to a like a smoke bar or a cigar bar.

Jacob Andra: Hold on one sec. So I wanna get back to the industry, but I'm very interested in this specific like side project.

Jason duPont: Good, brother.

Jacob Andra: Yeah, yeah. This is cool. So is it a custom GPT or is it something more involved or elaborate? Like how are you actually deploying?

Jason duPont: A bit more elaborate. Yeah, I went a little deeper on it. I have some memory tied to it, some RAGs, just different things built in to make it work and I'm constantly adding to it, so I'm adding body scans. I've wore Whoop, I don't know if I'd be able to show this to you during our thing, but familiar with Whoop and like the Oura Ring and all that? All those kind of things.

Jacob Andra: I do know the Oura Ring. I don't know Whoop.

Jason duPont: So they're just competitors. But I'll show you data. I don't know if you could see this though. Kind of. So I have a my day streak is 1,534 days. That's a lot of data, right? So I always want to keep adding data because it helps me personally get to my goals. So, for example, I have a tangent here. Here's a tangent. I have a goal that I want to lift a million pounds of weight. So move a million pounds of weight, not body weight, but in the gym in one session. I've been slowly working up to it. I just hit a quarter million the other day.

Jacob Andra: Nice.

Jason duPont: Through data because I'm tracking what I can do, I'm tracking how long it, 'cause those are brutal days, especially for someone like me. So my tracking is how long is my recovery gonna be? What do I do? What do I need to recover? What does my body work well with? What's gonna help me from in all the different levels? So it's funny 'cause the AI tells me all this.

Jacob Andra: Yeah, that. No, that's really cool. And there are, the reason I wanna press into this, even though this is a kind of personal side project for you, it's very relevant to our corporate clients and audience that, you know, the same thing you're saying can be applied to so many corporate use cases. So I'm very curious to the extent you're willing to share. So you're collecting all this data. You have some sort of a RAG functionality. Do you mind sharing the architecture or how you've built this? How you're collecting it, how you're storing it. How you're accessing it?

Jason duPont: I started with N8N and then just built it out on there.

Jacob Andra: Okay. So an orchestration.

Jason duPont: Oh, for sure, for sure. I mean, look, it's fully offline. I don't keep it all public, so it's all on my own thing, but you'd be surprised what you can really get into now and I can upgrade it as I go. The main piece of it was being able to look at large segments of data and be able to keep adding to it and then reference back. So just like DEXA scans like that, it can really see the differences. When it was early, like maybe a year ago, it might get confused on different timeframes and things like that. Those days are gone if you're building it right.

Jacob Andra: Yeah. And so what large language model, or are you using a combination of them?

Jason duPont: What I'll do is I'll switch according to, I don't care about the cost for me, like obviously in a business that matters and so you might go to the most recent version because of cost. Right? It might not even be that much difference too. You gotta weigh all those options. But for me, like with this, I don't care, so I'll just switch to whatever is working well with the most recent releases.

Jacob Andra: No, that's really cool. Thanks for thanks for being willing to take that tangent. Let's get back onto the mortgage track. So Tidal Wave, they're starting with the front end. So are we talking customer service chat bots, interfacing with customers? You know, what is the actual functionality you're deploying with?

Jason duPont: So when I say front, we're not starting with lead and we're not starting with CRM activities. What I would consider in that realm is that a lead could come from Facebook, a lead could come from a realtor, a lead could be a past client. They're all, I'm looping those together for this discussion. And then from there you have to have some sort of reach out. So we have things like, we're partnering with a company right now, Abound Voice AI. And there's a lot of them, right? So we had to really do our due diligence and see what we wanted for that. And I will tell you right now, me on a personal note, so I'm the top recruiter over at Nexa. I'm also the COO, right? So with that, as soon as I took on more of those COO duties, I had less time to meet with individual loan officers for recruiting. So what I do now is I have virtual assistants that are humans that help me get to a path where maybe they've seen things or done things before I talk to them. Right.

Well, the way that works, and this is kind of cool, so we have a Thursday Nexa call that talks all about Nexa. Every Thursday it's a Zoom. We do it on Thanksgiving, we do it all the time. I average, no, here's average, at least 10 people there every week that I've never talked to. That has never received a text message from me, at least in this campaign. Maybe they have in the past and that have not talked to a human that get there by outbound voice AI.

Jacob Andra: Okay.

Jason duPont: So that could be a referral. Somehow they get into my system right from there. They'll get to me eventually human right, but I want them to see about Nexa first. That's because of my time. It's a huge time saver because to get them there and that's the power of where outbound voice is now. So back to Tidal Wave. So that's not part of Tidal Wave, but that would be part of getting them into Tidal Wave system is those kind of things. It could be humans too, whatever. It gets them there.

And then the point of sale starts. The point of sale starts at 1003. When someone's ready to do an application, that is where Tidal Wave starts with us right now. They could change that in the future on their roadmap? But their roadmap extends the other direction before that direction. Our vision is to let CRMs connect to that and go that route.

Jacob Andra: Yeah. And so what exactly is Tidal Wave providing then? Can you walk me through the actual...

Jason duPont: Sure. So it's a full back and forth with the client. It can be text, it can be different things, but it's intuitive to help the client get there. Here would be an example. Let's say, you know, I've done loans forever. Say I'm at a restaurant talking to a client. I'm on my iPhone or what? I don't use the iPhone, but let's just say I'm on an iPhone, by the way. I have to convert due to our own or my coordinates, but you know, I'm gonna do it. So anyways, begrudgingly but let's say I'm at a restaurant. I'm taking a 1003. I'm taking an app just because I've done it forever and let's say I get 70% of it. 'Cause I don't want to go over relo while I'm there or something. Right? Well, I can just put that transcript in. The AI is gonna dissect it all and just start conversing with the client on the remaining questions that I missed. Or things that weren't there or things that maybe opened up because of the way I asked the question or what the answers were.

Jacob Andra: So it's an ingestion engine essentially to a context aware ingestion engine to get all the information needed from the customer.

Jason duPont: Yep. And I'm gonna plug my Mac in just real quick. Just gimme one second. Okay. There we go.

Jacob Andra: No worries. So it's ingesting a lot of necessary information from the customer, and then it's sort of forwarding that to the underwriting stage, right.

Jason duPont: So there's a big collection process, right? So you, and there's pieces on their roadmap that are not complete yet. Right? That's okay. We went into it knowing that. But there's everything from collection to connecting to different services. So you might collect, to connect to an Argyle or a TrueWork, it might pull a soft pull credit report, run a DU, it might upgrade to a hard pull at some point because of different things. So it's a whole ecosystem on finding what loan is best for the client, because a lot of times, or most of the time, the client's not gonna know. So slowly it's replacing that decision engine on what is best for the client.

Jacob Andra: Okay.

Jason duPont: VA, FHA, am I taking some non QM product? What am I doing here? And you can start knowing that. But and the vision is that it's trained to where, let's say, remember I'm in the TPO space, right? So I have different places I take loans, but let's say one lender has an overlay for something that they do and another lender doesn't. Just to make it real basic, well, the system, why waste time taking a loan to someone that has an overlay when I can bypass that? Or maybe income is calculated differently one lender versus another. So those are just kinda some high level things, or what if two lenders, what, sorry? What if the same lender, two underwriters calculate it different because of subjectivity? Then we have use cases that we can argue the income could still be good because seven times outta the last eight this went through this way and this one time it didn't. Let's talk about it, right? So now we have that all built in for that.

The vision of what we're doing though, Jacob, the whole vision is to get that one-touch close. We're very, very big on one-touch closes. We're not big on, you know, let's go back and forth five times on a loan. We're big on collecting everything, verifying everything, getting everything into place, and then getting that one touch close. What's that gonna do for the underwriter's life? How many more loans will they be able to do? The loans come in cleaner and cleaner and cleaner.

Jacob Andra: Yeah, so you've talked about a lot of aspects of the loan lifecycle. Everything from lead generation, getting leads into the system. You know, collecting information from the customer. There's a decision making process to route them to different lending options, obviously, as you said. Then there's underwriting, then there's compliance. So across, across, let's say underwriting, compliance and other aspects of the loan lifecycle, are there ways you guys are looking at AI implementations, you're already pursuing future roadmaps. How are you thinking about some of those areas?

Jason duPont: Oh yeah, for sure. I mean, we're looking at, look, there will be less processors in this world soon. There will be less underwriters, there will be less. I don't think that, like my vision is not that underwriters go away. My vision is that they'll be able to 10x what they're doing now. You know, first there'll be two x, three x, five x, but they'll be able to do a lot more because of AI.

Same thing with processors, right? So compliance piece will definitely be there. There will be loans that will never have compliance issues because it's caught up front by the AI, by just what's happened in the past, right? So a lot of these, a lot of these things, even the AI that we're working on, you're seeing where it just can't be trained by anyone or everyone. So you'll have like, you know, maybe a good word is like a contextualized in place. You'll have things in place to help really stay within certain guidelines and rules, especially in that space because these loans, not only are they regulated, but they need to meet in certain guidelines. They need to be profitable, they need to be done the right way.

But that, I mean, there's a lot that it's gonna, you gotta also realize loan industry, I don't know if fragment is the right word, but there's a lot of overlay with different things, and I don't mean overlay like where a lender won't do a certain condition or they're adding something to it. I mean, where redundancy almost, but not in a good way. There's just a lot of things that are done over and over and over and we're kind of antiquated still. I think that's why AI has gone other industries quicker than us because we, there's so many pieces. Plus we're regulated and there's a lot that has to happen for it to really get what I would call seamless. But in my vision the loan officer will always be involved to some degree.

Jacob Andra: No, I agree with that and I agree with your assessment that we'll just need fewer. Probably fewer loan officers, probably fewer underwriters, probably fewer processors, you know, all the different players in the ecosystem. We will still need them, just fewer of them, and they'll be doing far more volume with a lot of these technologies.

Jason duPont: Yeah, and underwriters, I think we could agree that the reason there'll be less is because AI will do more of the heavy lifting. Maybe that's a good way to put it. Loan officers on the other hand, I think they'll only be less. This is, again, see if you disagree on this. There'll be less loan officers, not because AI's doing more of the heavy lifting, even though it kind of is. It'll also be, or for me, more importantly, it'll be because certain loan officers are just gonna capture more market share.

Jacob Andra: Yeah, back, sorry, my phone. Back to that influencer dynamic you were talking about. You know, if loan officers are more like influencers, then obviously you get this Pareto principle where 20% or you know, some small percent capture the majority of business and everybody else is trying to pick up the rest.

Jason duPont: And we'll see. We'll see if it's full 80 20 or is it, you know, is it some crazy 80 20 of 80 20? Right? What will it really be? That's exciting and we'll see who uses it. Those that don't embrace though, they're just gonna feed everyone else.

Jacob Andra: Yeah. One of the things I like to educate the audience of this podcast on, and you know, as CEO of Talbot West, working with companies on AI, AI adoption, I'm always educating clients on is the different types of AI, machine learning technologies that, you know, AI is not just ChatGPT. AI is not just large language models. Obviously within the mortgage industry there are many different use cases for many different types of AI and machine learning. You wouldn't probably use a large language model to do like pattern recognition, matching, fraud detection. Right? That would be some kind of sophisticated machine learning algorithm or model. Right. And so maybe just talk a little bit about how you see different types of AI and machine learning technologies applied across the different aspects of the mortgage life cycle.

Jason duPont: Look, there's companies out there right now that are saying they're AI and they have AI, but they're still backing it up with humans. Not Tidal Wave, by the way, but I'm just saying there those are out there and it's because they're trying to catch up. So as AI comes into play, that will happen less and less and those things will happen less and less. But there are so many instances of where it, look, you have loan officers that are probably using more what you're talking about, like the large language models or maybe custom GPTs or maybe even a couple things beyond that where they're building some sort of like an N8N or Make.com or some of these other things they'll use. And there's so many out there now. I mean, they're coming up every single day. Right.

But I think the ones that are doing it, they're truly kind of pulling back, like Tidal Wave is pulling back and Nexa myself, Mike Kordes, a few others, we work very, very fast. The broker, the TPO channel, the non delegated, all that is a lot faster than the IMB space. Wrong or right. It's just how it is. They're a lot more corporate. We're a lot more people will use words like janky or they, right. We're, the all these words will come out. Tidal Wave is kind of in between. They want to move very fast, but they're gonna do it right. They have a very standardized release schedule. They're making sure that everything's in play. I love it. I still want to go faster. But I love it because they are doing it the right way, they're not of the mindset that we need to take months to get this one new thing out.

Jacob, you and I both know that AI is moving so fast that it is hard to keep up. I mean, tell me if I'm wrong on that. 'Cause it is hard for me to keep up and I'm in it every single day.

Jacob Andra: It's a very dynamic space. Absolutely.

Jason duPont: Yeah. Which makes it exciting.

Jacob Andra: I love it. Yeah. So is Tidal Wave largely based on large language models, does it have some other types of machine learning?

Jason duPont: Yeah. They're pulling it all back. I mean, look, will they reference them or were they, they gotta, of course, right? But they're pulling it all back. They're controlling the entire thing.

Jacob Andra: Yeah, I can imagine.

Jason duPont: To put those guardrails up.

Jacob Andra: Yeah, I can imagine that for the customer information intake portion, that's very well suited to a large language model, right.

Jason duPont: Yeah, I think that could be, but you start getting into dynamics real quick. I mean, think about like, you're gonna need some sort of OCR technology to start dealing with documents. You're gonna need to categorize properly. You're going to need to be able to index along the way. So there's a lot that goes into it. I think to really build like something that, like that homegrown would be difficult. I think there's other pieces that you could build homegrown, I'll give you an example. I think a loan officer could build a whole marketing thing fully done for them. And I think they could use things like 11 Labs and Hagen, all these different things where they get their avatar out there. And I think they could build it where you maybe tied it to WhatsApp or Telegram or something and you said, Hey, I wanna make a video on this and use my ethos.

And it all had encapsulated and you could write it all out. You could probably tie it in like a Claude or something like that, or whatever the model is for copywriting at the time. I think pretty much anyone could do or pay someone to do as long as they maybe had some technical skill or were young enough where it's just part of who they are. Right. I think if we ask the kids just coming outta college, or just graduating high school, I think that some of this is just kinda second nature.

Jacob Andra: Yeah. Are you keeping up much with the kind of trend of neurosymbolic AI? That's an interesting one I'm keeping my eye on, and I think there's a lot of applicability to the mortgage industry there.

Jason duPont: I think so. Probably not to the level you are. I'll just back real quick and then let's come back to that. So where I'm a big fan of and look people. AI is such a buzzword right now. It means so many different things. And so I oftentimes have to really say, okay, what do you mean by that? Because again, it's just such a buzzword. I don't even know another way to say it. Agentic AI is another buzzword, right? And it's becoming more and more popular and it probably has for a couple months now.

What I'm a big visionary of is just from my old marketing days. 'Cause I used to own a lead company for many, many years. Workflows. I think that that's very common to say. A lead comes in, a text goes out now, two minutes later, this happens, five minutes later, this happens, 10 minutes later, this happens. And I think AI was initially, well, not along the way, was built the same way where AI ran off workflows and but AI could make, you know, write the copy or research things. Then you added some memory to it and now it can reference things and say, okay, well this person likes basketball, so I'm gonna talk about the home team in their state or whatever. Right.

But just to define it, because again, people use different words. What I think is cool, and I think it's very, very close, especially when you put a bunch of agents in a hierarchy, where they're controlling each other or working with each other. Talking to each other is where you say, here's your task. The goal is this. Go get it done. That's not workflow based anymore. Is that agentic? By some definitions yes. By others. No. Right. But that's, and I know that's not what you're talking about, but I just wanna circle on that real quick because that is an AI that fascinates me is let's build these agents that work together. And so I have test agents that I've even built and I'm just nerdy enough to be able to do this stuff, right? And I have agents now that are building other agents without me telling them to. So I have like some crypto stuff and things that I do build other agents to achieve the goal. So anyways, but let me go back and we can come back to that later if you want to. But let's go back to what you were saying.

Jacob Andra: And I would love to revisit that 'cause that's another, I love some of the things you're exploring on the side in your personal life. But yeah, neurosymbolic. I mean that essentially, from what I can gather and some of the stuff that we're doing at Talbot West just means you are grounding a large language model or maybe an ensemble of large language models more in reality. You can use something like a knowledge graph or some other ontology that essentially gives it a set of guardrails. You know, a grounded source of truth. And so in the mortgage industry, I see, I could see that would be especially relevant because you don't want it going rogue. You don't want it fabricating stuff. You don't, you want it to stay very within specific boundaries and guidelines and so.

Kind of giving it a layer of ontological reality and do's and don'ts, and then pairing that with a large language model for that natural language interface seems like it could be very useful. And that's probably what Tidal Wave is doing. I hope that's what they're doing. It seems like that would be a really good idea for them to do. I'd love to explore further kind of what they're actually doing.

Jason duPont: Can, with that model, how do you choose which LLM you're using?

Jacob Andra: Well, I mean, really you could build any pairing or any, you know, we're big on kind of this open, interoperable, flexible sort of ensemble architecture where you might use a variety of different large language models and they're all referencing the same knowledge graph if needed. You could have different ontologies, different knowledge graphs that different large language models are querying about different aspects of the kind of reality. They're supposed to be operating within. There are a lot of ways you can structure it.

Jason duPont: I like it. I like it. I saw one company non-mortgage related that just kind of tied it. I think they built it too much to just one. Like, let's just say it was ChatGPT, for example, like they built it too much. But I think the, like what you're talking about, I like where you can vary it according to. You're just not tied to one thing because things change so fast.

Jacob Andra: Yeah, absolutely. Interoperability, I think and agility that way is really important. Let's get to that crypto and other use cases where you're building these agents. I, again, I wanna drill down into that 'cause it's so fascinating and it's the sort of thing where so many organizations could get so much value from doing that. So I'm assuming these agents are large language model based. Let me know if that's true or not. And then the second thing is what platform are you building them on to deploy agents that can create other agents, et cetera. How is that being orchestrated?

Jason duPont: Yeah, I'll get you a list because I toy with a lot of things. I mean there's Crew, there's there, there's so many different things and I don't want to give away my secret from my main one. 'Cause it, but I'll tell you offline on that, but they are mainly LLM based, but I do find, I'll just say in most cases, right, the high level of it is in my vision is that you're gonna see, you're gonna see people on that solopreneur list, right? The first one to a billion. I'm not on that list. I'm not going after that. That's not the play that I want, but I love that people are doing it and it's gonna happen. And what I see is as they build their own CEO agent. And then they have their hierarchy for their different things. And you see a, I see a lot where the most I see it in is marketing right now.

Like marketing agencies, I see that they have this person's over calling it a person, right? This agent's over design this, agent's over ads. This agent's over, organic social, this agent's over, right? Email. So but they work together to accomplish the main goal. And you'll see even where the CEO has is, this is kind of, and I had to build this for the crypto where the CEO has a like an auditor, but it's not the CEO, it's someone that they rely on so that when the reporting back happens, the auditor can say, does this make sense? Is this what I want? It, I don't wanna say it prevents hallucinations 'cause I don't really even like that word because I think you can get around that. But it does keep everything tight would be a good way to say it.

Jacob Andra: Yeah, yeah. At Talbot West, we've kind of pioneered this approach. We call it Cognitive Hive AI, but it's essentially an ensemble approach where you can have one, call it an agent, call it a, you know, whatever you want to call it. Agent is the current buzzword one module. This job is to audit or critique or hold another one accountable. You can set it up in a lot of different ways. And what you've described is a prime use case of that. And, you know, then you can expand obviously beyond large language models to kind of bypass their inherent weaknesses, but still take advantage of their strengths by pairing them with other types of machine learning or, you know, more I guess you could say more deterministic type models or ground them in reality with different ontologies.

That neurosymbolic approach. There's a lot of ways you can take it, but I think for a kind of a DIY person who wants to get pretty into the tech and use some of these platforms you're using that, you know, doing it with a set of large language models initially is a great way to get your feet wet with that.

Jason duPont: Sure, for sure. And look, there's, if someone wants to do it themselves, there are so many YouTube videos and that's a whole nother rabbit hole because you're gonna find out these new platforms that come out, the all the no code stuff, all the light code stuff. That's my word by the way. But all those kind of things that are out there to do this on and if you're not doing it like on an enterprise level, it can be done very cost effective. If someone's doing some of the things that I'm talking about with like a whole hierarchy and a CEO and all that and you would do an enterprise there, there would be some cost to it. There would be because of everything that goes into it. Right. But you can do it the other way.

I'll give you another thing on that. So we talked about that I kind of have a set it and forget it recruiting thing that I haven't even looked at for months. Right. But it just happens. And then my, I have four virtual assistants that help me full-time, that are humans, right? They're human, amazing. With that, on the outbound voice AI that helps bring people to the Thursday Zoom call. Those calls are recorded and those calls are looked at by another agent to determine how the call went. The prompting for future calls is then modified based on the execution, so it's constantly making itself better.

Jacob Andra: So that kind of recursive feedback loop, you know, we call reinforcement learning, obviously. And so that's cool. You've built that into your process.

Jason duPont: I oversimplified that there's probably like six things in there that have to work together to do that. But essentially that's what it does and what I've been working on, it's funny 'cause that was built first for my AI stuff. I know there's services out there that do this for humans, but the next thing is we do it for my four virtual assistants too.

Jacob Andra: Yeah. That's awesome. Well, anything else you wanna say about the mortgage industry, the future of it? Oh, let's have you say that, tell that one anecdote about the one-touch close, you're getting closer and closer, and you were telling me when we had our kind of intro conversation that, you know, you had achieved an important milestone. Let's hear about that really fast.

Jason duPont: Which one was that? What you mean? You mean the one that I personally did before we did AI or?

Jacob Andra: No, I think it was involving AI. I don't remember the details, but it was like you had achieved a very, like truncated close. It was really impressive.

Jason duPont: No, I mean, we've done so much. I'm not sure which one it was. I know, I will tell you this. So this is, there's no AI involved in this, and so this is, the vision is that I have achieved with my team a 54 minute from submission to clear to close. So what that means is...

Jacob Andra: That's the one I was referring to.

Jason duPont: The time that it's submitted to the lender to the time that it was clear to close was 54 minutes. Not conditional loan approval, but actual clear to close. So the next step is go there. Now, on that, we had to wait out three days or certain things we had to do because of the way the rules work. Because of that, because we can do that with humans, our vision backwards is how do we do that with AI, right? And so every single loan that goes through, we're looking at everything. We're looking at the conditional loan approvals, we're looking at everything. And why is that there? Why is that still there? How do we solve that? And that's what helps us get there quicker and quicker and quicker. And so as we partner with different companies, as we do different things. Those are the solutions we're working for and looking for. I'm sorry.

And it's interesting because as we do this, we have so many companies hit us up. Can we partner with you? Can we partner with you? And every now and then we'll find one. You know what? That makes sense. That would fill in that gap. Because Tidal Wave has a goal and we love Tidal Wave, it's not just like a contractual partnership. There's actually a, we like them. I mean, it's really cool that they're amazing partners. Diane, who leads the thing over there, amazing individual, amazing, right? But along the way we have to get to this clear to close. So when we have other companies hit us up, we're like, okay, would that fit in temporarily? And then how do we deal with that? And, you know, is this long term, is this short term, but to back it up, that is the goal is how do we get that on scale? And it's all about scale.

Jacob Andra: Yeah, that's great. And you know, I'd love an intro to Diane. It sounds like a really great company.

Jason duPont: Oh yeah. Yeah. She's amazing. And she's done a lot of cool things in her life, and she's formed an amazing team. And I think that's a big piece of it is those that you surround yourself with make a huge difference in your life. And from a work perspective, she surrounded herself with some amazing, amazing people.

Jacob Andra: Awesome. Well, hey, it's been a great conversation. Thanks so much for your time.

Jason duPont: Yeah. Yeah. Appreciate you Jacob. Let me know if you need anything else.

Jacob Andra: All right.