Financial Services
Senior Mortgage Underwriting System
Four specialist AI agents assess a mortgage at the same time, with a person stepping in on every contested case.
01The problem
Underwriting a mortgage often takes three to five days. A big reason is that four separate checks (credit, income, savings and the property itself) wait in line behind each other. The work also carries real rules about fair lending and personal data that cannot be skipped to save time.
02What it does
The system gives each of the four checks to its own specialist agent. A supervising agent hands out the work and gathers the results. A critic agent then reviews the combined picture before a final recommendation: approve, approve with conditions, or decline.
03How it works
Personal details are removed before any AI sees the file. A bias check looks for signs that a decision leans on something it should not. The agents look up the lender’s written policy rather than relying on memory. Any contested file is sent to a human underwriter, every time.
04What came out of it
On three test applications the system reached the expected outcome each time: one approval, one conditional approval and one decline. Running the checks side by side points to a cycle of hours rather than days. Building it also showed me the limit of letting AI make the decision, which is why my later Databricks version keeps the decision in fixed rules.
Built with
Want the technical detail?
The code and the full technical write up are on GitHub.