Underwriters were spending most of each file re-reading familiar documents. A copilot that reads first, cites its sources, and flags what has changed now supports every decision.
The challenge
A mid-sized lender’s underwriting team was turning files around in days rather than hours, and most of that time was spent reading: bank statements, employment letters, appraisals, and the same forms in a hundred layouts. Automation had been attempted twice and rejected both times, because no one was willing to rely on an answer they could not verify.
Our approach
We began with evaluation rather than model selection. Two hundred historical files, graded by senior underwriters, became the test set every version had to pass. Retrieval anchored each answer to the page it came from, the interface placed the citation beside the claim, and every override an underwriter made fed the next evaluation run.
What we built
- Document ingestion and extraction across the full form library
- Retrieval-grounded summaries with page-level citations
- An evaluation harness of graded historical files, run on every change
- An audit trail of suggestions, overrides, and outcomes
- Deployment within the lender’s own cloud tenancy, with no data leaving it
What changed
- Underwriters begin from a cited summary rather than from page one
- Every suggestion can be traced to a source the compliance function can inspect
- The lender can change models or vendors without rebuilding trust from the beginning
This engagement is representative of the work Loonwise undertakes; client names and specifics are shared in conversation.
Contact
From ambition to working software.
Whether you are accelerating AI adoption, modernizing a critical system, or building the capabilities to innovate faster, we bring the consultants who scope the work and the engineers who build it—the same people, from the first conversation. A senior engineer responds within one business day.
Get in touch