From Paper to Platforms: Rebuilding Mortgage Servicing with Andrew Wang (CEO and Co-Founder of Valon)

all right well scott welcome to our podcast again i'm so excited to be here with you again not our back backdrop that we had at salt flats but it is pretty it is not julie the same backdrop because that one was about as fun as it gets and yes i do want to affectionately call it fintech burning man going forward because that really was julie but you think about that we did we did that one up you know fintech burning man now we're here in miami so i don't know what we're going to do for the next one like maybe we need to do something like the top of snowbird like in february or march something like that we'll do some crazy podcasts at the top of the tram you know we can ski down after i don't know how we're going to top these last two though we can go up to the north pole and see what santa thinks about fintech i like that idea we could do a little maybe santa's choosing AI to speed up the process of delivering presents and like figuring out what kid wants what, making sure things don't get messed up.

There you go. We could do a whole podcast on Santa AI. I love that. Let's do that.

Well, you know, someone that knows a lot about AI though, is our guest today, Andrew, who is a co-founder of Valen, a mortgage servicing company. And I've done some work with you guys, talked to your team and learned more about your company. And, you know, I almost think of you as an AI company in many aspects versus a fintech company. So for anyone that's not familiar, can you tell us a little bit more about what you guys are doing and what the idea behind Valen was?

Yeah. First of all, thanks for having me. So what we do is we build regulated OSs for the mortgage servicing space. And when people think about what mortgage servicing is, they think about it as, hey, there's this business of collecting payments from homeowners.

You're obviously dealing with all of the intricacies of the regulatory policy associated with managing that mortgage. And for Valen, what we've sort of re-envisioned that entire process as is you can actually build a fundamental infrastructure core that does all of the underlying details on its own. And then you can orchestrate AI workflows on top of that that allows the operators at the end of the day to actually provide the type of service that they want for their consumers.

And so it's a pretty interesting dynamic where we fundamentally change the way that people think about mortgage servicing from one that is very antiquated and old using effectively pre-internet systems to one where they're now able to describe what they want to do for their customers in human language and then get that orchestrated down to the infrastructure layer as an effective set of processes and procedures that are compliant with the rules and regulations. Yeah. Well, Scott, I assume you also own a house, right?

I do own a house. Yes. Although I would like to have something out on the salt flats after our beginning of this, but yeah, so I do own a home. Yes.

So I have two mortgages technically now, not a great market in Austin. So we're renting out the one and we just bought a new house, but talk to me a little bit about like how my servicing experience, I think you are now working with the company that I have my servicing agreement with. My loan got sold off to. That's amazing by the way.

Amazing. There you go. Congratulations. I haven't experienced it yet, but supposedly in the coming year, I'm supposed to start experiencing.

What am I going to notice from my end user experience? How is it going to improve? What's going to be different? So maybe the first thing I'll start off with is the mortgage space.

And I'm talking about both origination and servicing hasn't really been innovated on over the last 50 years. And a lot of that has to do with the fact that, look, at the end of the day, the mortgage business and how mortgage loans are created and sort of how they flow throughout the economy is a very government-oriented and backed process. And that's an amazing thing for a lot of different reasons. People are talking about it these days more actively because of Trump's recent, or I guess maybe it was Pulte or Trump's, recent proposal around the 50-year mortgage.

But the more important thing to sort of succinctly realize is that, hey, there's a lot of regulations, there's a lot of rules around it, and there's a lot of implications if you try to change it. Now, what that also means, though, is because of this myriad of rules and regulations, again, that makes it hard to actually do what you want to do, it means that the servicers have to use a lot of paper and processes. Most of these guys work on hundreds of systems to get anything done.

Meaning, you're like, hey, I want to know, did you pay my insurance bill? And they don't actually really know, right? They know they maybe made a ledger entry. They know that someone in the Treasury Department maybe moved a bunch of money but they can really tell you what going on right not to the level of the detail that we i think as a you know fintech industry sort of almost naturally assume whereas with valon what we've been able to do as i mentioned just earlier is build that infrastructure layer really really cleanly so that you can orchestrate everything on top so it's a very long window way for me to basically say what you will be able to expect is when you want something when you're asking something, all of it actually can be automated because before Valen, what these guys would not get comfortable with is the fact that it would be done compliantly, even if they use AI.

Yeah, yeah, yeah. With Valen, if you're like, hey, I need to know X information, they will have the infrastructure supported by us to say, I know exactly the answer. I can provide it to you in a way that complies with all the regulations that I feel really, you know, required to follow. And so you'll get that instantaneous response.

Now, God forbid you go through some sort of situation where you, well, either that, you need a loss of mitigation, you need assistance, or alternatively, maybe you want to change the terms of your loans through what's called a recast process. Before, you actually just submit a bunch of paperwork, you mail it in, you wait whatever, 30 days for them to turn it around. Now, you can just send it in and it comes back to you immediately. Yeah, it's amazing.

It's interesting, Andrew, Julie, what I think about mortgage systems, and you know, I'm inside and out, Andrew, I think of them very similar. It's like core banking systems. So my favorite joke is like, you know, listen, and I'll say this affectionately. I was born in the 70s.

I love the music from the 70s. I like a lot of things about the 70s. I just don't necessarily want my mortgage system or my core banking infrastructure, those types of things to be built on technology from the 70s. And I think, Andrew, it's so interesting about what you guys have built is you've got all of these different AI tools now that are really hopefully going to allow these streamlined processes to actually really take place and not have these just archaic, painful, kind of manual processes that I think are so consistent with what you typically see in the mortgage business.

Yeah, I'll even expand upon that before. And to your point, like the systems that were built in 70s were done with the constraints. It's the guys who built it, built it the way that they had to build it, given the circumstances. Yeah.

And like we look at them, we're like, honestly, when we look at them, sometimes we think to ourselves, wow, they really like did the best that they could with what they had. But the reality, and then as you know, people who are obviously very interested and focused on the developments of AI, there's only so much you can do when there's a literal mainframe system in the back closet. And you're like, great, I have all this AI. Can the AI walk over to the closet and open it and like, you know, pull information?

It's like, no, not really. You actually need to build a vertically integrated piece of infrastructure so that you can pull the information, do all the different stuff. Yeah, no, Andrew, you touched on something that I'll just back over to Julie, but it does make me laugh because you look back and she's like, oh, why did they do that? You're like, hey, listen, let's not forget.

We went to the moon in the 60s, right? Think of the technology from the 60s. We actually pulled some pretty cool stuff off. So I agree with you.

It's like you look back, it's like, hey, there's some really smart guys that built some stuff. But now you look 50 years later and you're like, wow, look at the stuff that you're working on, I think it's just incredible to see how you can really optimize that entire space. Yeah. Well, I was born in the 90s and I don't know that I would want that technology also powering a bunch of financial services at this point either.

But you mentioned this is like 50, 60 years ago. What do you think, since you've seen some of the evolution and things that you're hoping to do, what do you think the industry looks like even just 10 years from now into the future? So what I think is unbelievably powerful about the current set of LLM tools is that it allows you to do a variety of really hard to do tasks previously by machines, right? Because fundamentally, what people have really tried to do is break down the processes and things that they need to do as a financial industry into things that are deterministic rules, like things that we know exactly heuristically how to move from point A to point B, and things that are not so obvious.

And that not so obvious camp actually is a very long tail. It can be something as simple as, hey, what is the actual dollar amount from this bank statement? And then if you take the next step, you're like, well, does this line item in the bank statement, is that an income item or is that someone wiring money in? You kind of have to make a determination.

And a lot of those like fuzzy decision making problems, right? Previously were done by humans because, you know, machines aren't very good at it. And they also want to be able to get comfortable with explainability, asking questions outside the system, like reasoning, like, oh, maybe I actually need to go ask the homeowner before I actually make the next step. And so there's a lot of these different parts of it that fundamentally were difficult to do before.

Whereas today, if you, again, have the right infrastructure, the long tail can actually be solved quite nicely with LLMs And the other nice part about it is that the base LLMs the way that they have been built are trained on a corpus of knowledge that allows them to reason sort of logically like a human would, which means that if you have the data associated with how you should do these discrete tasks and steps, you can actually quickly either fine-tune or, you know, train the models to actually do those specific tasks based on the data that they have.

So what that means is that you now are in a world where let's say there's a thousand, three thousand random small tasks that you have to do for this process. Before you're not going to train all these models on each and every individual task. Now you can actually use the LLMs to go solve them. And what that, again, in a very long window way, means is that you can actually have a world where all of these long complicated processes are automated.

And what we as humans and customers experience in the industry is effectively a fully automated process. It is actually quite possible in the next, or when I say quite possible, I would have bet very strongly that in the next 10 years, from the mortgage industry's perspective, both on originations and in servicing, it should be 99% automated. Well, one question I have for you, you mentioned data a lot in that answer. And one of the things I've noticed using some of my personal finance tools, like bringing everything into one app and whatnot, is things on credit card statements, et cetera, can be labeled differently depending on you know, did I pay for that Uber with my credit card?

Was it Uber Eats or was it a tip for the ride? How do you guys think about the data that you're getting internally and making sure that's super clean? And is that something that you work on internally or do you work with one of the other fintech providers that's here like Spade, for instance, that really works on cleaning up that data? On our side, we think of the world as you're either in the infrastructure business or maybe another way I like to characterize it.

And by the way, I'm not the only one who says this. Actually, there's a recent paper that the guys at Rivet put out. There's like the token factories, and then there's the token injectors and model companies, right? Our job is to be the best token vendor.

What that means is effectively, I am trying to produce the cleanest, the best, the most reconciled data for the LLMs to act on. Because we all know LLMs, you know, still hallucinate even the recent models still have some degree of that right and the better data the better context and the better tooling that we can provide to these tools well the better the outcomes and the better decisions will be so we spend so much of our time thinking and whenever we talk to our customers we tell them we're in the business of actually modeling the world we want to be able to represent the truth of the world and what's interesting is that the world changes your systems will try to follow as best as possible and to minimize that drift.

But the more accurate and the more closely tied to the real world we are, the better decision making we can give, the better and more consistent information we can provide to consumers. And so we will work with companies like Spade to triangulate that data. We will go buy or integrate with as many data sources or we'll even go as far as, as an example, going back to my insurance payment example, we'll literally have the check tracked when we send it to the insurance company. So we know actually every state of that check, right?

Do we literally print the check? Do we send it out the door? Is it in the mail? Has it been checked?

Like all these different states we can provide and that flows up in terms of information to the LLM to then make the best decision or provide the right information possible. Andrew, you mentioned something that I thought was very interesting. You said how, you know, LLMs can hallucinate. I had this very conversation yesterday and I was actually right by the bus that everybody's spray painting back there.

And so it was so interesting to have this conversation about like how LLMs can, you know, hallucinate, et cetera. And I was like, for a second there, I'm like, wait a minute, am I having this conversation or is this from all the fumes from the, from the paint over there by the bus? But no, it's, it's a very interesting, interesting topic, but I thought it was like, it was, it was really funny to have that right while I was standing next to the, all the fumes, Julie, from the bus.

All the fumes. And you know, something that as you're talking and telling me about what you guys are doing for mortgage servicing, it It sounds very applicable to other complicated areas as well. Where do you think that you guys could work in the future outside of mortgages? We've actually looked.

So we started servicing unsuitary consumer loans. Okay. We've also looked at solar and some other categories. So I would say anything in the loan servicing category is an obvious thing, given that we've done one of the most difficult, if not the most difficult, consumer loan servicing.

But fundamentally, what we've built and what we've really focused on is actually multi-party ledgers, right? Where we're constantly doing reconciliation with all the different counterparties and constituents. And what that means is that any business where there a lot of money movements across different people and you have to constantly think about who owes what person what and think about it from the perspective of not just cash movements but receivables that have been accrued or transactions that have occurred.

When you have that plus a ton of regulation, that's when we're doing the best because that's the type of tricky, messy problem that it's difficult to represent what's going on in the world. And the historical systems have not been able to do that well. So we very much are against working or looking at spaces where it's too simple because we don't think we have any true value added advantage. Whereas the more complex, the more messy it is to represent the world, that's where we are very excited to work at.

Hey Andrew, along those lines on the regulation side, because it's a very interesting topic of saying, because the mortgage space is a highly regulated business. And you think of AI and is your model able to take regulations and be able to adjust to that? Right now, the pendulum of regulation has kind of swung this direction, maybe after the midterm elections or in 2028, they'll swing back the other way. How do you think about AI and regulation in the mortgage space?

I'm super curious to get your thoughts. This is a really great topic. And I'll use an analogy. I'll first say my single statement and hopefully pithy way of thinking about it, and then I'll give you my analogy here.

the pity way is everybody in mortgage has gone through the great financial crisis and they've gotten questioned by the regulators they've been testifying in front of Congress and so they sort of conceptually think about it as the quote-unquote testify in front of Congress test they always ask themselves what would I say if something went wrong and I'm asked to testify in Congress why I did a certain action and so that's like a deathly paralyzing fear that these guys have now what does that mean and actually I had this conversation with the folks at FICO probably a decade ago when there was this boom, if you guys remember, around gradient boosting, random foras, different sort of models and machine learning techniques to basically predict whether or not someone was gonna default.

And I asked the guys at FICO, I said, hey guys, it's not like you don't know these techniques exist out there. Why don't you guys use this? Why do you still use logistic regression? And they looked at me and they basically said, Laura, it's really, really simple.

It's not that we don't know this is a more powerful technique. It's about explainability. It's about that the actual modeling technique that used it so simple, still powerful enough, but so simple that if you were to be asked in front of Congress, hey, why did you do X? You can very easily say this parameter literally increases your FICO score.

This parameter now decreases your FICO score. We don't have to talk about it. There's no like weird intersection between, hey, you're like from this demographic and you're like this income and this job. No, no, no.

It's very clearly someone who makes more money, better FICO. Someone who makes less money, lower FICO generally, like some version of that, right? And that is really, really important because then you feel comfortable going for a product and explaining them what's going on. And so the way I think about compliance and AI and really the financial services industry is that, yes, you could use LLMs to do a bunch of the processes, whatever else, but as much as possible, you want to think about building your infrastructure and a product in a where the key components of what you will be questioned on will be deterministic and explainable.

So say that said differently. Instead of saying, I'm going to let the LLM decide whether someone should get a loan, I will actually use the LLM to convert my policy of how I want to give loans into a deterministic workflow that I've reviewed, and then I'm going to put a human through that. Because then it's like, I've actually checked it, right? So another example of this is, would you rather use an LLM to write a Python function to calculate something that you're trying to determine, or would you just ask the LLM?

And the answer is for the LLM, it can sometimes just make some stuff up. But if it wrote the function, you can actually review the function, get comfortable with the function. And it's explainable at that point. It's fully explainable.

You literally know line by line how that works. I love that. Amazing. Well, Andrew, thank you so much for joining us on this episode.

This was great. I learned a lot about how AI can change all of this. This was so quick, and thank you for having me. Thank you.

Scott, we've got to do this again. North Pole with Santa. Let's do North Pole with Santa. I really like that idea, Julie.

We'll have to do that because, again, this one and the one before are going to be hard to talk, but I think I like that North Pole idea. I've always wanted to know what Santa thought about FinTech, so this is it. Exactly. And now he's, I'm sure, implementing AI as we speak.

Exactly. Exactly. All right. Well, thank you, Andrew.

I appreciate it. Thank you all. We'll see you next time.