Mortgage Intelligence

Mortgage Intelligence Platform

The intelligence layer underneath every application.

The platform turns mortgage documents and data into a source-linked mortgage representation, connects that representation to governed policy and deterministic tools, and supports reasoning across the lifecycle.

Architecture

Applications on one intelligence layer, on one mortgage context, on source evidence.

Mortgage Intelligence Platform architectureFive applications, QC, Underwriting, Closing, Delivery and Servicing, sit on one Mortgage Intelligence layer made of six pillars: Data, Knowledge, Reasoning, Tools, Evidence and Evaluation. The layer rests on canonical mortgage context, which rests on source evidence. Only QC is available for evaluation.APPLICATIONSQCAvailable for evaluationUnderwritingPlannedClosingPlannedDeliveryPlannedServicingResearchMORTGAGE INTELLIGENCEUse models for judgment where judgment belongs. Use deterministic systems where truth can be computed.DataKnowledgeReasoningToolsEvidenceEvaluationMortgage contextcanonical, source-linked mortgage data: borrowers, property, income, assets, liabilities, terms, program, historySource evidencedocuments with page and position, delivered data, lifecycle history, reviewer actions

Read from the top: five applications consume one intelligence layer. The layer is six pillars. The pillars work on a canonical, source-linked representation of the mortgage, and every fact in that representation traces to source evidence. Only QC is available for evaluation; the other applications carry their status on the diagram and on their pages.


The principle

Use models for judgment where judgment belongs. Use deterministic systems where truth can be computed.

It would be simpler to say the model reads and the rules decide. That is accurate for quality control today, and too narrow for what the platform is built to do. The division that survives the next several product generations is by kind of work, not by component.

MODELS
perceive, interpret, synthesize and reason
KNOWLEDGE
establishes the policy and context that apply
TOOLS
perform deterministic calculations
RULES
enforce deterministic requirements
PEOPLE
resolve judgment, ambiguity and exceptions
EVIDENCE
preserves provenance and auditability

The product does not persist or expose hidden model reasoning. It exposes a structured decision trace: evidence, policy, calculation, tool outputs, assumptions, missing data and disposition.


Six pillars

What the layer is made of

01

Data

A normalized, source-linked representation of mortgage facts and lifecycle context.

02

Knowledge

Effective-dated policy, program requirements, rules, exceptions and overlays.

03

Reasoning

Domain intelligence connecting facts across documents, policy and lifecycle context.

04

Tools

Deterministic calculations, validation engines and workflow operations where correctness should not depend on generated prose.

05

Evidence

Traceability from material conclusions back to source data, documents, rules, calculations and reviewer actions.

06

Evaluation

Mortgage-specific testing of correctness, applicability, evidence and professional work products.


Perception and data

Reading is a capability anyone can buy. The representation is not.

A vision-language model classifies the documents, locates fields and handles the variation of real-world files. That runs on managed infrastructure rather than a model of our own, because document reading improves across the whole industry every quarter and is not where a mortgage platform should spend its differentiation.

What the platform builds from that reading is the part that cannot be bought: a canonical mortgage record in which borrower, property, income, asset and liability facts from many documents become one structured representation, with the source of every value kept, the conflicts between sources carried forward rather than resolved silently, and derived values shown with the formula that produced them.

The honest present: today a general-purpose vision-language model from an established provider reads the documents, and a versioned, deterministic rule engine produces the findings, with a person resolving the exception. Customer documents are not training data. How data is handled.


Knowledge

Effective-dated policy, encoded and versioned

Not a checklist held in someone's head. Each control is a stored rule, bound to a loan program, with a semantic version, the person who published it and a lifecycle status. A finding is bound to the version that produced it.

Mortgage policy changes. Selling guides are updated, program requirements shift, compliance interpretations move. A review performed in March was performed under March's rules, and by September those rules may read differently. If a system cannot say which version produced a finding, every historical review silently re-anchors to today's rule set.

The knowledge layer encodes the GSE selling guides, FHA, VA and USDA program requirements, and the federal compliance frameworks including TRID, ATR/QM, HOEPA, HMDA, RESPA and Regulation B, as rules selected by loan program. The control set as it applies to quality control.


Reasoning

The part above reading, and what it has to get better at

Recognizing an income figure is the reading layer. Explaining why qualifying income differs between two programs is the domain layer. This is the specific list of what mortgage reasoning has to do, and each row marked work in progress is exactly that.

TODAY
Locate a value on a page
A model finds monthly income on a paystub and reports where. This works, improves industry-wide, and is not a differentiator.
IN PROGRESS
Know which value the program actually wants
Qualifying income is a calculation, not a field. A 24-month average, a year-to-date annualization and a declining-income treatment can all be defensible for the same borrower, and the right one differs by program.
IN PROGRESS
Tell a legitimate difference from a defect
A paystub, a W-2 and a 1040 cover different periods. They are not expected to match. Telling an ordinary difference from a real problem is what separates a finding from noise.
IN PROGRESS
Know what should be in the file and is not
Describing what a file contains is reading. Knowing a required document is absent depends on program, property, occupancy and jurisdiction. A system with no applicable rule reports nothing, and nothing reads as clean.
IN PROGRESS
Explain an exception well enough to act on
A reviewer facing a 203(k) binder or a DSCR file does not need a flag. They need the evidence assembled, the requirement named and the open question stated precisely.

Today is the first row plus a versioned rule engine and a human reviewer. We will publish evaluations on representative, unseen mortgage cases as they exist: expert-reviewed accuracy, exception handling and the usefulness of the explanation.


Tools and rules

Deterministic where truth can be computed

Calculations run as code, with the formula and inputs shown. Hard requirements run as rules that contain no model, so the rule applied, its logic and its version do not vary between runs.

It would be convenient to say that because the rules are deterministic, model error cannot affect the outcome. That is not true, and an experienced reviewer will see through it immediately. If extraction reads monthly income as $12,000 when the paystub says $7,000, the debt-to-income calculation is entirely deterministic, and entirely wrong. The rule did not change. The conclusion did.

What is genuinely fixed is narrower: given the same inputs and the same rule version, the same result follows. So the design goal is not infallibility. It is that a misread is visible and correctable before it becomes a decision, which is why every value is shown with the document it came from, and why the system produces validation findings while a reviewer reaches the disposition.


Evidence and human workflow

Anyone can produce a finding. The question is whether it survives being questioned.

Six months after a file is cleared, someone asks why. A finding that cannot answer that question is an opinion with a timestamp.

A numeric check that passes, an evidence question that still needs a person, and everything kept with both. FIGURE Anatomy of a finding Four sources, four periods, one comparison, and why the difference needs a person. RULE INC-014 v2.3.0 · QUALIFYING INCOME MUST RECONCILE WITHIN 5% ACROSS SOURCESSOURCEAS STATEDMONTHLYHOW DERIVEDPaystub 03/2026p.1 · YTD box$26,820 YTD over 3.0 months$8,940 / moYTD ÷ months elapsedW-2 2025p.1 · box 1$106,500 for 12 months$8,875 / moannual ÷ 12URLA 1003p.2 · 1c$9,200 stated monthly$9,200 / moas stated by borrower1008 Summaryp.1 · field 21$9,200 used to qualify$9,200 / mounderwriter’s figureSpread between lowest and highest monthly figure: 3.7% inside the 5% threshold.Not a defect. Flagged for review because the 1008 uses the stated figure, not the derived one.The reviewer decides: accept the underwriter’s basis, or request a written income calculation. the system records which,by whom, and when. The finding is an exception to investigate: the disposition belongs to the reviewer. Illustrative, from a synthetic file. Threshold and rule identifier are examples.
A numeric check that passes, an evidence question that still needs a person, and everything kept with both.

Every value carries the source document it was read from, and every finding the numbered rule version. Which rules ran, which were skipped and against which loan program is recorded per run, so a disputed answer is traceable to a specific value on a specific page. The reviewer's disposition is captured with the deciding reviewer recorded. That is a design position, not a gap, and it is not on a roadmap to be removed.

Unresolved judgment and exceptions are escalated to a named person rather than silently invented. Queues, assignment, escalation and service levels around that reviewer are on the roadmap, and the QC page says where they stand.


Evaluation

Every release measured on mortgage work

Domain-specific tests of correctness, applicability, evidence and professional work products, on representative mortgage cases. Generic AI benchmarks are not the measure.

Extraction fidelity
Document and field values read correctly, with the page they came from.
Reconciliation
Cross-document consistency: what should agree, and whether it does.
Policy applicability
The right requirement, in the version that applied on the relevant date.
Calculation correctness
Deterministic math that matches the professional work product.
Missing-data recognition
Knowing what should be in the file and is not.
Defect detection
Precision and recall against outcomes an expert reviewer already knows.
Evidence completeness
Whether a finding carries everything needed to defend it later.
Exception handling
An exception stated well enough for a reviewer to act on it.
Reviewer acceptance
Whether the people who do this work accept the finding as useful.
Cost per successful case
What a correct, complete, defensible result costs to produce.

Results are published with the method, the document set and a comparison anyone can reproduce, or they are not published. The evaluation harness itself stays internal. Our research approach.


Modular by design

API-first, which makes the pieces separable

Relevant if you want part of this rather than all of it, or want to put it behind your own front door.

The application layer consumes the same core services used by the platform itself. Extraction and synthesis, the versioned rule engine, and the evidence and lineage layer are distinct services rather than one indivisible product. That separation is designed to support integration and modular deployment over time.

Availability of individual services is scoped per engagement and should not be inferred from the architecture alone. A firm that already has a reviewer workflow and wants the rule engine and evidence record behind it is a conversation we can have; so is a partner wanting to deliver quality control to their own clients. Neither is a shelf product today, and we would scope either as an engagement rather than a purchase.

We will take a finding apart.

A working session against your form mix and rule set, on a synthetic file or de-identified files of your own.