Latest / Chrisman Commentary - Daily Mortgage News / 8.11.26 Conference Chatter; Redwerk's Konstantin Klyagin on AI Governance; Bond Directionality

Transcript

8.11.26 Conference Chatter; Redwerk's Konstantin Klyagin on AI Governance; Bond Directionality · Chrisman Commentary - Daily Mortgage News

• 0:02
  Speaker: Welcome to the Chrisman Commentary Daily Mortgage News Podcast.
• 0:06
  Speaker: I'm your host, Robbie Chrisman.
• 0:08
  Speaker: Topics on today's episode include Chatter in the Halls, a Western Secondary, Why Rates Edged Higher Yesterday in a Quiet Trading Session, in my interview with Redworks, Konstantin Klyagin on how AI and mortgage lending must be governed as an entire decision-making system.
• 0:24
  Speaker: Here, take a listen, do a little preview.
• 0:28
  Speaker 1: When we think about what lenders or originators or banks need before their AI deployments reach a regulator, what do they need?
• 0:39
  Speaker 2: Discipline, and part of that discipline is uh documentation.
• 0:43
  Speaker 2: So they need a clear pipeline and it must be documented.
• 0:47
  Speaker 2: Because if you don't know what you're automating and you expect a certain outcome, it will not happen.
• 0:52
  Speaker 2: AI works like it amplifies garbage.
• 0:55
  Speaker 2: So if garbage goes in, then you have more garbage out coming out.
• 1:01
  Speaker 2: And uh so, first of all, the pipeline that you are automating must be clear.
• 1:05
  Speaker 2: You need to be able to reproduce it manually before you automate it.
• 1:09
  Speaker 2: You need to know what happens on every step.
• 1:11
  Speaker 2: Then you need the escalation logic, and then you need to know when it uh shouldn't answer.
• 1:19
  Speaker 2: But the clear pipeline is key.
• 1:21
  Speaker 2: Without clear pipeline, you cannot do a good uh you cannot create a good AI solution.
• 1:30
  Speaker: Thanks to this week's podcast sponsor, Optimal Blue.
• 1:33
  Speaker: Did you know Optimal Blue's profitability center unifies pricing, hedge performance, pipeline activity, profitability, and market intelligence into one personalized dashboard, giving mortgage lenders faster, more complete insights to make better capital markets decisions.
• 1:47
  Speaker: To learn more, visit optimalblue.com.
• 1:50
  Speaker: What are folks talking about here at the Western Secondary?
• 1:56
  Speaker: Small things and big things.
• 1:58
  Speaker: There's a brand spanking new private mortgage insurance company named Anza.
• 2:03
  Speaker: Lenders are doing interesting things to help communities.
• 2:06
  Speaker: For example, Fairway Independent raised nearly half a million dollars to bring highly trained service dogs to veterans and first responders at no cost.
• 2:15
  Speaker: UWM filed a lawsuit in federal court regarding the Two Harbor Cross-Country deal, to nobody's surprise, although $500 million is attention grabbing.
• 2:26
  Speaker: China unleashed $28 trillion, yeah, with the T capital markets to challenge the United States.
• 2:32
  Speaker: We've heard about profitability, but volume is also important.
• 2:36
  Speaker: According to Kiranos, proprietary application index, July 2026 funded mortgage volume was flat for the year, but was down 7% month over month.
• 2:46
  Speaker: The average 30-year conforming retail funded rate in July was 6.42%, eight basis points higher than June 2026, and 33 basis points lower than the same month last year.
• 3:00
  Speaker: 30-year mortgage rates don't necessarily track 30-year Treasury bonds, but long-end rates shot up sharply last month, with the curve bare steepening as an oil shock and the seventh consecutive interest rate hold by the Federal Reserve pushed inflation risk into forward expectations rather than the policy path.
• 3:18
  Speaker: Great markets are likely to keep testing the Fed's resolve, though the pressure on agency mortgage-backed securities seems to be volatility driven.
• 3:27
  Speaker: Agency mortgage-backed security issuance remained resilient in July, with gross issuance totaling $110.9 billion, the fifth consecutive month above $100 billion and extending a 25-month streak of year-over-year supply growth, even as activity eased modestly from June.
• 3:44
  Speaker: The housing market continues to cool in an orderly fashion, characterized by low turnover and challenging affordability.
• 3:52
  Speaker: Home price appreciation is slowing, but affordability remains strained as higher prices and elevated mortgage rates push household mortgage debt to record levels and keep borrower leverage elevated.
• 4:03
  Speaker: Debt-to-income ratios have improved slightly from recent peaks, but remain well above pre-pandemic levels and are a stronger indicator of mortgage risk than LTB ratios, which can quickly deteriorate if home prices reverse.
• 4:17
  Speaker: With household equity at record highs and refinance incentives gradually returning, cash-out refinancing activity has begun to recover.
• 4:25
  Speaker: While today's mortgage market is fundamentally healthier than before the financial crisis, lenders should remain focused on borrowers' ability to service debt rather than value of the collateral servicing it.
• 4:38
  Speaker: Recent regional disparities and home price developments appear to be largely a function of inventory, which is normalized, suggesting these regional differences might shrink in coming months.
• 4:49
  Speaker: With the understanding that the Fed must allow data to drive its decisions and communicate that any policy shift is a response to economic conditions rather than political pressure.
• 4:58
  Speaker: U.S.
• 4:59
  Speaker: agency mortgage-backed securities started August strongly, delivering the best weekly excess returns since early April.
• 5:06
  Speaker: Declining volatility and resilient prepayment expectations have supported the sector.
• 5:10
  Speaker: With Fannie Mae 20-year and higher duration coupons leading performance.
• 5:15
  Speaker: Spreads remain attractive relative to treasuries and modestly cheap to investment grade corporates, particularly in Ginny Mae 30-year and Fannie Mae 15-year securities, while select lower pay-up and mid-coupon vintage pools offer relative value, although fallout from the ongoing Middle East conflict warrants continued caution.
• 5:32
  Speaker: For today's interview, I wanted to welcome to the show Red Work's Constantine Kleiden to talk about how AI and mortgage lending must be governed as an entire decision-making system, requiring rigorous testing of data, integrations, permissions, escalation, edge cases, and audit trails.
• 5:51
  Speaker: He's founder of Red Work and overall a tech entrepreneur.
• 5:56
  Speaker: He has decades of experience in software development, and his career has spanned roles as a software developer, tech lead, project manager, and technical writer in both startups and large international companies, leading to the establishment of his IT services business.
• 6:12
  Speaker 1: It seems like there is a gap between what people want technology to do or what they think it can do and what it's actually doing.
• 6:23
  Speaker 1: And AI is obviously advancing very rapidly, but it seems like companies are shipping things before they've actually validated it.
• 6:33
  Speaker 1: Is that what you're seeing from your end when it comes to lending institutions, deploying technology?
• 6:40
  Speaker 1: And if so, kind of why why does that gap exist?
• 6:45
  Speaker 2: Of course.
• 6:45
  Speaker 2: So with the with the widespread of uh vibe coding and AI assistant software development, we do see a lot of uh acceleration in the shipment of features and generally in like release cycles and in creating new software products, but AI only accelerates the coding part.
• 7:07
  Speaker 2: And the rest of the software development lifecycle requires discipline in terms of uh specification, documentation, and also test coverage.
• 7:16
  Speaker 2: Because if you don't know what you are automating and what you are applying AI to, then uh the outcome is not expected.
• 7:25
  Speaker 2: While AI makes things faster, it also requires the standard discipline that's been around for many decades uh related to the rest of the software and lifecycle.
• 7:39
  Speaker 1: Are companies realizing this problem?
• 7:42
  Speaker 1: Are are they aware, or are they are in from your experience, is it something that uh they have to get called out on after the fact, and then they will make changes?
• 7:56
  Speaker 2: Well, it depends on the business.
• 7:58
  Speaker 2: I cannot generalize.
• 8:00
  Speaker 2: So some uh well, I saw some proper applications of uh AI in software development.
• 8:06
  Speaker 2: For example, we have a customer who prototypes on their end, then they show us the feature they want to have uh in their products, and then we rebuild it properly, and then we release it to production.
• 8:19
  Speaker 2: They know that the prototype they build is not production ready.
• 8:24
  Speaker 2: Everyone uses vibe coding nowadays, and Genai for uh for coding is uh pretty much in every uh business already, but uh those companies that get it right, they uh they have technical background and they have like professional software developers and architects behind the uh prompts.
• 8:47
  Speaker 2: So if it's done properly, there is architecture, infrastructure, security, aware, coding, and then you can ship world-class uh products.
• 8:58
  Speaker 2: Otherwise, it's prototyping, which is which AI is also great for if that answers the question.
• 9:04
  Speaker 1: Yeah, do you do you have thoughts on the best ways to test and validate models before shipping them?
• 9:12
  Speaker 2: Well, yes, uh, you should be aware that they don't follow the regular deterministic logic and test them accordingly.
• 9:21
  Speaker 2: But what's also important is not uh, and I need to stress this first, is that it's not only the model that you have to test.
• 9:29
  Speaker 2: There is uh a bunch of other pieces of the puzzle that a uh modern-day AI-enabled solution has that need to be tested alongside the models, and that is the data, like if the context and the data uh that is used in this decision making is right, then the APIs that it's pulling, and is it actually pulling the APIs or hallucinating?
• 9:54
  Speaker 2: You need to test the APIs, you need to test the permissions because uh with the different classes of users, it may need or may not need access to certain uh data or actions, then the escalation logic is very important because uh the uh solution needs to know when to involve a human.
• 10:16
  Speaker 2: Uh, then logs are important, it has to log everything it does, and also the edge cases, because people and businesses they tend to test the happy path.
• 10:28
  Speaker 2: And if you test the edge cases, then you are way safer.
• 10:33
  Speaker 2: Because, especially in mortgage, especially in fintech, any uh incorrect decision leads to uh like real problems, like money loss, for example.
• 10:43
  Speaker 2: And uh there is something that those mistakes may be really expensive.
• 10:50
  Speaker 1: How can companies produce better audit trails?
• 10:53
  Speaker 1: Seems like there's a common problem out there that a lot of companies can't necessarily pinpoint problems because of this.
• 11:02
  Speaker 2: Regular logs, like regular deterministic logs with uh logging every fork of an algorithm of a deterministic logic, are not enough with the AI and with the agentix softwares.
• 11:15
  Speaker 2: Because uh, well, you need to always log the context, the uh decision making, uh, the outcome, the handoff.
• 11:24
  Speaker 2: There are much more points to log.
• 11:26
  Speaker 2: Because if you cannot reproduce, if your audit trail doesn't let you reproduce every step in your decision making, then uh you can get problems with regulators, for example.
• 11:38
  Speaker 2: They are moving slowly, slower than uh than the innovators, which means which is good, which is a good thing because there is more window to establish like uh internal control, but they can be after you, and if something uh if something uh goes wrong, you need to have an audit trail, you need to have an extensive log of every decision, every step, every input, and every output that you have in your AI solution.
• 12:08
  Speaker 2: Because if you don't don't own your AI agent, then your AI agent owns you.
• 12:16
  Speaker 1: Oh, that's a good line, right?
• 12:18
  Speaker 1: So, how can we get to better validation of non-deterministic AI outputs?
• 12:27
  Speaker 1: And maybe that gets into what compliance teams should expect from their QA stack here with technology evolving so rapidly.
• 12:35
  Speaker 2: In Fuinberg, uh, it's one of my agencies, we have established a certain protocol for uh genetic AI testing.
• 12:43
  Speaker 2: And uh the most important part, but it's valid for any type of software, it's not just uh testing the happy path.
• 12:50
  Speaker 2: So you need to try different like contradictory inputs because people are not always consistent, right?
• 12:56
  Speaker 2: Consider the simplest application, an AI-enabled chatbot, and the inputs you can be getting from your customers may not be consistent at all.
• 13:05
  Speaker 2: And uh, we are not always always rational, you know, as humans.
• 13:10
  Speaker 2: And you need to test that, you need to feed it garbage, you need to you know try it in extreme situations.
• 13:16
  Speaker 2: But the layers of testing that we have on our list is we first test the uh factual accuracy.
• 13:24
  Speaker 2: So if the data that is operates, uh that uh it's operating is up to date, then we test the context, yeah, we make sure that all the details, all the relevant details are taken into account.
• 13:37
  Speaker 2: Uh, then also a very important part is the boundaries.
• 13:40
  Speaker 2: We test boundaries.
• 13:42
  Speaker 2: Does the bot refuse or the AI agent knows when not to answer, when not to take action, when it has to escalate to a human, for example, or just do nothing, because sometimes doing nothing is less expensive than doing the wrong thing.
• 13:58
  Speaker 2: Uh then also you need to test escalation.
• 14:01
  Speaker 2: So uh if the bot if the uh agent knows when to involve a person.
• 14:06
  Speaker 2: Lastly, you have to test auditability.
• 14:09
  Speaker 2: Can an institution reconstruct what actually happens and when certain things happen happen, when uh when let's say a mistake popped up and influenced all the rest of decisions made downstream.
• 14:24
  Speaker 2: Uh so if you can reconstruct it, then you own your solution.
• 14:28
  Speaker 2: If not, it owns you.
• 14:29
  Speaker 1: I would be remiss if I didn't ask you what you're working on.
• 14:32
  Speaker 1: Can you talk about the latest uh latest products and uh what you feel like the company solves?
• 14:39
  Speaker 2: Yes, yes, of course.
• 14:40
  Speaker 2: Uh so uh just to remind uh you and your listeners, I run two agencies.
• 14:45
  Speaker 2: One is QA Work, which focuses on uh quality assurance, and the other one is Red Work, which is uh a software development agency.
• 14:54
  Speaker 2: If you are asking about QA Work specifically, then uh we do uh out of the most recent and the most exciting AI-enabled solutions that we've been testing is Granola AI.
• 15:08
  Speaker 2: It's an uh it's an uh it's an AI note-taking uh app uh valued at $1.5 billion.
• 15:16
  Speaker 2: And they came to us exactly for discipline because they know, and they are like they are very strong technically, they know that they cannot be just constantly shipping features.
• 15:27
  Speaker 2: Uh, what they also needed is a strong partner for uh software quality assurance.
• 15:33
  Speaker 2: And this is how we have integrated ourselves into their release cycle.
• 15:38
  Speaker 2: And every release gets tested, gets test coverage.
• 15:42
  Speaker 2: We create automated tests for what we can, and what cannot be automated, we test manually before every release because they know that the mistakes and bugs in production are very dangerous because they result into customer churn and ultimately into you know less margin of their business.
• 16:00
  Speaker 2: Something more relevant to the uh subject of your uh show is uh our fintech solutions that we've been working with historically.
• 16:09
  Speaker 2: For example, uh it works on the African market, it's a chat app with encrypted messaging, but also with payments.
• 16:19
  Speaker 2: So you can chat and send payments.
• 16:21
  Speaker 2: Then there is also Zazu, uh, it's the finance man uh management app, uh, which is a MasterCard principal member.
• 16:30
  Speaker 2: Then there is Union 54, also African.
• 16:33
  Speaker 2: They uh they appeared on TechCrunch from Business Insider.
• 16:38
  Speaker 2: And another example is uh Iconomy.
• 16:42
  Speaker 2: Uh it's a uh UK-based crypto asset management platform.
• 16:47
  Speaker 2: So these are our FinTech uh customers.
• 16:51
  Speaker 1: You're a busy guy for sure.
• 16:53
  Speaker 1: I appreciate you making the time for me, and uh really enjoyed the discussion today.
• 16:57
  Speaker 1: So thank you very much.
• 16:58
  Speaker 2: Thank you.
• 16:58
  Speaker 2: Thank you for having me.
• 17:03
  Speaker: The bond market sold off to open the week as renewed optimism around sustained U.S.
• 17:07
  Speaker: pressure on Iran pushed crude oil prices higher, lifting five-year and longer yields to one week highs while the two-year reversed Friday's rally.
• 17:14
  Speaker: Despite repeated optimism from U.S.
• 17:17
  Speaker: officials, negotiations over reopening the Strait of Hormuz have produced no tangible progress, leaving oil prices rising as geopolitical uncertainty has persisted.
• 17:26
  Speaker: As a reminder, July's unexpectedly weak jobs report, marked by the first negative payroll print of the cycle, sharp downward revision, slowing wages, and falling labor, forced participation, undermined the case for a September rate hike, leaving this week's inflation data as the key test for the Fed's next move.
• 17:44
  Speaker: CPI is expected to have remained temperate in July, with pressures becoming less widespread and more concentrated in specific sectors.
• 17:52
  Speaker: Today's economic calendar kicked off with July's NFIB Small Business Optimism Index in at 99.8, its highest level since August 2025.
• 18:01
  Speaker: Later today brings Red Book same-store sales, July existing home sales, which are expected in at 4.07 million, and results from a $58 billion three-year Treasury note auction.
• 18:12
  Speaker: We began Tuesday with agency MBS prices slightly worse down or down from Monday's close, the two-year yielding 4.24, and the 10-year yielding 4.71 after closing yesterday at 4.70%.
• 18:23
  Speaker: Let's wrap up with a joke and some housekeeping.
• 18:29
  Speaker: Here are some texting codes for seniors.
• 18:32
  Speaker: Obviously, young people have their codes for texting, but now seniors have their own codes.
• 18:38
  Speaker: ATD means at the doctor's.
• 18:41
  Speaker: And WAITT.
• 19:01
  Speaker: Who am I talking to?
• 19:03
  Speaker: Thanks again to Optimal Blue for sponsoring this week's podcasts.
• 19:11
  Speaker: Optimal Blue's Profitability Center unifies pricing, hedge performance, pipeline activity, profitability, and market intelligence into one personalized dashboard, giving mortgage lenders faster, more complete insights to make better capital markets decisions.
• 19:24
  Speaker: To learn more, visit OptimalBlue.com.