In this week’s AI Update, we unpack the buzz around Agentic AI—intelligent agents designed to handle tasks autonomously, from booking travel to processing a mortgage loan. While the promise is compelling, the pitfalls are real: without careful design, agent-based systems can quickly devolve into “spaghetti code,” creating inefficiencies instead of eliminating them. Join David Lykken and Pavan Agarwal as they break down what agentic AI really means, how it applies to the mortgage industry, and why thoughtful engineering—not just smarter models—will determine its success.
[David] Listeners, it’s time for another AI update! and of course, I have as my guest Pavan Agarwal of Celligence, also better known as Angel AI. Pavan, good to have you back on.
[Pavan] Good to be back, David so I got some interesting stuff to talk about. Everyone’s talking, obviously, trying to figure out how to put AI into their organization, and there’s all this stuff called agentic AI. Some people have heard that.
[David] Would you explain what agentic AI is real quickly for those that go? We hear that but I’m not sure everyone understands that You’re the AI expert. What is agentic?
[Pavan] Yeah, so it’s just using AI to actually perform specific tasks.
[David] Okay.
[Pavan] So, instead of, like, you go to ChatGPT and say write me an essay for my school homework and it’ll generate something. Now, you can say basically, you plug in ChatGPT into your process and it’ll start doing specific things, whether it’s booking a hotel or sending a response to customer service requests or things like that. But it’s a complex thing to do. So, let me give you an example. So let’s say you’re traveling to California. You go to California for a conference and you tell the AI book me a hotel in Venice. Okay, next thing you know, it’s booked you a hotel in Venice, Italy.
[David] Ah, lacking specificity.
[Pavan] Right, but it should understand that, hey, if it had a proper, complete context, if it understood David Lykken, it would understand that David doesn’t fly to Venice Italy often. It flies to Venice, California is more likely, and it should have been looked in your calendar and it should have said oh yeah, hey, there’s an entry in his calendar that says there’s a conference in Venice, California, so that means he needs to go to Venice, California. You see how complicated this becomes, and so the whole point of Agentic AI is to write an AI that you stuff it with enough context and you stuff it with enough rules, right, so it could do these things autonomously for you.
[David] That’s really good.
[Pavan] Right. And so what happens is like with frequently within all kinds of engineering, when something new like this comes out, it sounds great, oh, we’re going to have AI do it. But what ultimately happens is people start replacing spaghetti code written in Java or C or something and replace it with spaghetti code written in AI prompts right. So they’ll write an agent to do something. And then there’ll be edge cases that they haven’t thought about, and now they got this big, long list of prompts trying to confine it, and what happens then is when it gets so complicated, right. Then they get it to work somehow, but it starts taking up, as the prompts get bigger, as the request gets more complicated, it starts taking up more and more GPU power. So there’s a protocol out there called MCP – Model Context Protocol, which is open source protocol, which is supposed to solve this problem, and it does. It’s an amazing protocol that solves the problem if you follow it properly. So, now I just talked about the simple example of booking a hotel in Venice for a conference. That’s actually a very simple agent compared to approve this loan or fund this loan. Here’s a loan application and I want you to fund it. That’s what we do in the mortgage business right. Yep, David Lykken wants to buy a home in Venice. Okay, and fund this transaction. Essentially is the command.
[David] Right.
[Pavan] Right. So it has to understand everything about David Lykken. It has to understand that it’s Venice, california, not Venice, Italy, right, it has to know your bank accounts. It has to know where to wire it from it has to understand the CD and all this. So this is the challenge that’s in front of someone that wants to write an AI-based loan or any kind of banking system, you have to have agents who do all of these things, and if you don’t design this right, you end up with spaghetti agents and you have no idea what. Because of hallucinations, you don’t know what they could end up doing.
[David] Yeah, and I want to again, we’re educating our audience on this. When you say agent, you’re talking about a piece of technology that will write code, and that’s when you’re talking about that, am I you want to expand on that?
[Pavan] About that? Am I my agent? I’m talking about a piece of technology that is AI. That is essentially a piece of ai software that will do a specific task for you right,
[David] Exactly right, all right, good, good, good. And there’s a growing number of agents that we’re seeing on the marketplace that are really having an impact, it looks like they have a promise. I don’t know if they’re having the full impact, but they are being used. But then it’s the unintended consequences of those agents when it starts writing, as you say, spaghetti code, which means it’s kind of a mess.
[Pavan] What I’m talking about is spaghetti code. Is the developers developing agents without fully understanding what the edge cases are and how it could get out of control? So, you know, we kind of shifted the problem from writing big software systems with a lot of code you know, java or something else to big AI system with a lot of AI prompts, which is a form of code, a different kind of code, right. So you went from it seemed like it’s always in computer science it always seems like the next new technology will solve all the problems from the past, right, all the engineering problems from the past. But will solve some problems, but it creates a new set of problems. A new set of problems, yeah, and always there’s a lesson to take away here. Always the problems stem from bad engineering. You can’t, in other words, you can’t replace the need for quality design, quality architecture, quality engineering, no matter what technology you use, and the promise is always hey, this new technology is going to eliminate the need for us to have really high quality engineers and scientists. No, that doesn’t work that way. It doesn’t work that way.
[David] And that’s what you have developed at Angel AI. You have developed something that has the right balance. Could you just give us contrast what we’re talking about at Agentic AI to what you have built with Angel AI?
[Pavan] Yeah, so you have to give like the example of booking your hotel in Venice. So it has to understand. For that AI, that agent, to work correctly, has to understand the whole context, has to understand everything about David Lykkenn, so that when you give it a command hey, book my hotel in Venice it knows what to do, okay, so the fundamental architecture is that it has a high level understanding of not just the mortgage business, it also has a high level understanding of that particular loan file that it’s working on. So it works out. It solves a problem in a different way. It starts by what we call top-down. It starts by understanding the big picture and then when you say, here’s David Lykken wants to buy a house, then it goes to understanding David Lykken and then his income and all that, and then it finally goes through all these layers of understanding and then finally executes. Okay. So, whereas instead of the focus on let’s execute, let’s wire the money, let’s close that loan and pray it got the rest of it right, we have an AI that focuses on complete end-to-end understanding and then starts drilling in deeper.
[David] So the way you’ve built Angel AI is you have AI and you also have humans involved in the process, and a lot of people say, well, it’s not AI if it’s got humans involved. Well, you’re addressing some of the issues that you’re talking about right here by having, like the Angelistas and so many of the others involved to make for a better outcome.
[Pavan] Yeah, so there’s a company called Scale AI and it just got valued for like over $20 billion. I saw that it’s crazy, right, and it’s really an army. They have tens of thousands of employees. It’s an army of humans, okay, and they’re valued so much because what they do is they’re the data science company. So all the big models out there use them to organize and plan and map the data Interesting. So the real value of AI, what differentiates one AI system from another, is the data and the data science. Okay, so it’s going to turn out because Scale AI did the data science for all these other companies right, and when the dust settles, we’re going to see that they’re actually worth more than everyone else because they have the complete, exhaustive knowledge across all the partners that they work for right, and that’s where the really secret sauce is in the data science, okay, and the other guys, they’re just hardware companies at the end of the day. I mean they have, you know, big, massive arrays of GPUs but the intelligence is with scale. To answer your question, you can’t avoid this process of data science and data organization and tagging the data and preparing it for AI. We call that, there’s a field of academic study called augmented intelligence.
[David] Augmented intelligence
[Pavan] you can Google that it’s been a long-standing research for decades, and the idea is AI and human working together and so many of our patents and many of our designs are around how you do that in real time? Okay. So, in other words, what scale AI does is it scrubs it and it gets a whole bunch of data, prepares and tags it and then gives it back to, say, Gemini, and then they use their tagged data from scale and then run it through the GPUs to train their model. Okay, but we’re in the mortgage business and things change really fast. We need to do this in real time. So when something comes at us from out of left field new regulation or whatever we need to put in the AI and the AI needs to figure out as best it can, loop in a human assistant, make any corrections and update it and give their response back, and all that needs to happen in a minute.
[David] So, what’s so interesting about it? That’s why you’ve secured over a billion-dollar investment out of Japan to invest in mortgages. And this is just the tip of the iceberg. Folks, if you have not seen that press release, I’m going to put it in the show notes. But Pavan has landed a billion dollar investment where Japanese investors are willing to put a billion dollars in the US housing market, provided it’s run through Angel AI. Because of this partnership of human and AI to get it right. They don’t have confidence in the other way. We’ve been doing mortgages for all these years and that’s why it’s really exciting. Pavan, I want to wrap this up real quickly because I’m trying to keep these two within a short period of time, but it’s so much fun to get you on. We’re going to be having Jennifer on in our next episode of this. I’m really excited about it. In our next episode of this I’m really excited about it we’re talking about the pitfalls of using the large language models, the LLMs out there, to relying on that to making decisions, and you said she’s an expert, so I’m really looking forward to having her on. Tell us a little bit about Jennifer and tee this up for next week’s podcast.
[Pavan] Oh, she’s a genius.
[David] Oh, she is, oh, no, she is.
[Pavan] She’s amazing. Oh, she’s a genius. Oh she is. Oh no she is. I don’t know what her IQ score is, but my guess I’d be surprised if it’s less than 150. Yeah, probably close to 200. So she’s our CTO and COO. She’s the main engineer, main designer behind Angel AI and she’s the one who makes, put all the pieces together, make sure it all connects and works together.
[David] Yeah, well, I can’t wait to have her on next week. So listeners, get tuned. We’re going to be talking about that. Pavan, thanks so much for coming on. I’m really enjoying it. We’re getting a lot of plays on this AI update weekly update that we’re doing each and every week, and so it’s grateful, I’m grateful for you to come on and share your knowledge and wisdom. Appreciate it, friend.
[Pavan] Thank you, David Bye.
Important Links
- Press release about Japanese investor bringing AI powered innovation
- 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.