On the Ledger

The episode surfaces a few genuinely useful ideas - converting LLM outputs into deterministic workflows for compliance, and the 'token vendor' framing - but a significant portion of the 19 minutes is consumed by location banter, Santa jokes, and the host's personal mortgage anecdotes, diluting the substantive per-minute yield.

“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”

“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”

The 'testify in front of Congress' mental model for compliance decisions and the concrete analogy of asking an LLM to write a Python function vs. asking it directly are genuinely fresh framings; the rest largely recycles standard AI-in-fintech narratives about legacy mainframes and data quality.

“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”

“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?”

Andrew Wang is a credible practitioner who has actually built and operates a regulated mortgage servicer, and his FICO explainability anecdote shows genuine domain depth; however, the short runtime and conversational format prevent him from fully demonstrating the operational scale of what Valon has achieved.

“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?”

“We've actually looked. So we started servicing unsuitary consumer loans. Okay. We've also looked at solar and some other categories”

There are a handful of concrete specifics - gradient boosting vs. logistic regression at FICO, tracking every state of an insurance check, Spade as a data partner, a 99% automation claim - but the episode contains no revenue figures, customer counts, timelines, or case-study data to substantiate the broader claims.

“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?”

“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”

The hosts ask five broadly relevant questions and one decent data-specific follow-up mentioning Spade, but there is no substantive pushback, no challenging of the bold 99% automation claim, and the conversation is repeatedly interrupted by personal anecdotes and jokes that kill momentum.

“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?”

“How do you think about AI and regulation in the mortgage space? I'm super curious to get your thoughts.”