AI Document Management Portal: Designing Intelligent Package Review for Mortgage Underwriters
Overview
Designed the Document Manager, an AI-powered module within Freddie Mac Multifamily's MyOptigo platform, enabling underwriters to review loan submission packages with greater speed, accuracy, and confidence. The tool uses AI to extract key property data from submitted documents, including org charts, appraisal reports, and financial statements, then surfaces discrepancies and accuracy signals so underwriters can validate without manual cross-referencing.
The project spanned the full package check-in workflow: from the moment a lender submits documents, through AI extraction and completeness checking, to structured feedback back to the lender.
The Problem
The existing package check-in process was entirely manual and email-driven. When a lender submitted a package, the assigned underwriter received an Outlook notification, then spent 30 minutes to over an hour manually reviewing each document for completeness and data accuracy before drafting feedback. For full underwriting reviews, the process could stretch to a month, with multiple back-and-forth cycles between UW and lender.
Org charts posed a particular challenge: complex, multi-modal documents that AI struggled to process reliably, yet critical to establishing borrower structure. High-risk documents like appraisals required a "second eye" but humans could easily miss discrepancies buried across multiple reports. There was no centralized place to see whether extracted values agreed across documents, or to understand where the AI's confidence was low.
Research
Conducted interviews and workflow observations with underwriters and analysts across the Freddie Mac Multifamily team. Mapped the full package check-in lifecycle, from lender submission through UW sign-off, identifying where time was lost and where errors were most likely to slip through.
Key findings: underwriters wanted role-based document views, not one-size-fits-all lists; cross-document inconsistencies (like conflicting zip codes across Appraisal, PCA, and ESA reports) were a major source of rework; and AI extraction outputs needed to be editable and explainable before underwriters would act on them. Security concerns around borrower structure information also shaped how org chart data could be surfaced and stored.
Design Solution
Designed the Document Manager around two core views: Property Summary and Document Comparison. Property Summary aggregates AI-extracted data from all submitted documents into a single table, showing each data point alongside an accuracy score, number of source documents, and a flag for inconsistencies. Underwriters can see at a glance where the AI is confident and where values conflict across sources.
Document Comparison lets underwriters select specific documents and compare extracted values side-by-side, defaulting to their role's most relevant document set but switchable via dropdown. Highlighted rows surface discrepancies instantly, eliminating the need to open documents individually to find conflicts. A side rail allows access to a compact version of the Document Manager from anywhere in MyOptigo without losing context in the main workflow.
The feedback workflow was redesigned to let underwriters flag specific documents and send structured feedback directly to lenders from within the portal, replacing ad hoc Outlook emails with traceable, document-linked requests.
Testing and Iteration
Conducted iterative design reviews with underwriters, production analysts, and product owners. Prototyped the Property Summary accuracy scoring to test how underwriters interpreted confidence signals, refining the visual language to distinguish AI-low-confidence from data-genuinely-missing. Tested the Document Comparison default view with different roles to ensure the right documents surfaced first without requiring manual configuration.
Iterated on the side rail placement to ensure the Document Manager remained accessible without interrupting the existing loan review workflow in MyOptigo.
Outcome
Reduced manual cross-referencing time for underwriters through AI-assisted extraction and side-by-side comparison. Structured feedback workflow replaced email, creating a traceable audit trail between UW and lender. Accuracy scoring gave underwriters a clear signal for where to focus human attention rather than reviewing every extracted value equally.
The Document Manager established a reusable pattern for AI extraction with human validation across Freddie Mac's MyOptigo platform, applicable beyond the initial package check-in use case.
Key Learnings
AI confidence needs a visual language. Underwriters wouldn't act on AI outputs without understanding where the system was certain versus where it was guessing. Accuracy scores and inconsistency flags became the trust layer.
Role-based defaults matter. Surfacing the right documents first by role reduced cognitive load significantly. The same comparison table felt overwhelming in a one-size-fits-all view and manageable when pre-filtered by context.
Replace the email, not just supplement it. Structured in-portal feedback only worked when it fully replaced the old Outlook workflow. Partial adoption would have created two parallel systems and more confusion.