I built an underwriting tool using AI - Claude. Trained off historical loan performance and loss analysis.
I built an underwriting tool using AI - Claude
trained off historical loan performance and loss analysis
but the guiding decision maker is income performance after loan loss
what did the loan, and others that looked like it, actually produce, even when it went bad
Trying to figure out the real use case
training for underwriters so they can see patterns from past loan loss and past income generation
thought guiding while you underwrite, pulling up similar profiles and outcomes
first net catch on applications to save time before deep diving
Main thing I keep coming back to
how do we see past the ratios with a tool
DTI, score, all the usual stuff, useful but it hides the story
LSCI, “see what no one else sees”
how do you build that without making it a black box - or is this better because it is real performance? I have a hard time accepting that, but worth thinking about
I built in score codes
little prompts that force one step further
what to look at next
why this file is different than it looks
where the risk actually shows up
where the opportunity hides too
Still early
still testing it against old deals and comparing to current loans being underwritten to see where it differs from underwriting decisions
still letting it evolve
Let’s see where it goes