E3

Mortgage intelligence

Mortgage intelligence,
built for the depth of the industry.

Our aim is to bring expert mortgage understanding to every stage of the loan lifecycle. We are building specialised intelligence across mortgage documents, policies, calculations and exceptions — with quality control as the first application, running on real files today.

Bring a recent QC file See how a finding is built

What we are building

Mortgage runs on documents. Understanding them is the whole problem.

Quality control is the first application, not the boundary of the ambition.

Every consequential decision in this industry rests on a stack of documents that no single person ever reads end to end. Whether to lend. Whether to buy the loan. What to service, and on what terms. What sits inside a security. Whether a file will survive an examination in three years. Each of those is a judgement about what a pile of paper actually says — and each is made, today, by someone reading a fraction of it.

That is the gap. Not a workflow to automate, but a layer the industry has never had: a system that reads mortgage documents, understands what they assert, knows the policy that applies, and can prove the answer to anyone who asks later. Get that right and it serves the originator, the aggregator, the investor, the servicer and the auditor — because they are all asking questions of the same documents, and all being answered today by a person with a checklist.

Owning that layer needs a model built for this domain, not a general one pointed at it. General models read a loan file without reasoning about one. Mortgage work is thin in general training data relative to how specific it is — a 203(k) case binder, a twelve-month bank-statement income calculation, the way a rescanned addendum breaks a layout — and the gap shows up as plausible answers that a reviewer has to catch. So we are building a mortgage-domain model — and the rule engine is what earns the right to it, because it produces exactly the signal that is otherwise unavailable: which rule fired, what a reviewer did with it, and why.

How that improves the system, stated so a buyer can check it against the data section below. Nothing is learned from your borrower documents by default. What improves E3 for everyone is our own work on the rules — a reviewer override tells us a rule was wrong, too broad or badly explained, and the fix is a new rule version, reviewed and published, not a silent model update. Your policy configuration and your overrides stay in your tenant and are applied to your loans. For anything to become reusable knowledge outside it, three things have to be true: a mortgage expert has validated the conclusion, it has been tested against a held-out set rather than the files that produced it, and you have agreed to it in writing under a separate agreement. Absent all three, it stays yours.

What we will not do is claim any of it before it is true. Today E3 runs a general-purpose vision-language model behind a versioned rule engine, and a person resolves the exception. That is the honest description of the present. The destination is a system the whole chain relies on — and "industry-leading" becomes publishable only alongside a stated benchmark and a comparison anyone can reproduce.

The domain model

Reading a value and understanding a loan are different problems

Document reading is now a capability anyone can buy, and we say so on the platform page. What cannot be bought is the mortgage reasoning that sits on top of it. That is what a domain model is for, and this is the specific list of what it has to get better at.

TODAY: READING
Locate a value on a page
A general vision-language model finds monthly income on a paystub and reports where it found it. This works, it is improving industry-wide, and it is not a differentiator.
THE HARD PART
Know which value the programme actually wants
Qualifying income is a calculation, not a field. A 24-month average, a year-to-date annualisation and a declining-income treatment can all be defensible for the same borrower — and the right one differs between an agency loan and a bank-statement programme. Choosing correctly, and explaining the choice, is domain reasoning.
THE HARD PART
Tell a legitimate difference from a defect
A paystub, a W-2 and a 1040 report different periods and different definitions. They are not expected to match. Knowing which differences are ordinary and which are evidence of a problem is the judgement that separates a finding from noise.
THE HARD PART
Know what should be in the file and is not
Describing what a file contains is a reading task. Determining that a required document is absent depends on programme, property, occupancy and jurisdiction. A system with no applicable rule reports nothing, and nothing reads as clean.
THE HARD PART
Explain an exception well enough to act on
A reviewer facing a 203(k) case binder or a DSCR file does not need a flag. They need the relevant evidence assembled, the applicable requirement named, and the unresolved question stated precisely. That is the output we are building toward.

Every row above marked the hard part is work in progress, and the honest description of 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. Those are the measures that would substantiate a claim of leadership. Coverage counts are not.

The first application

Quality control, running on real files today

The rest of this page is what E3 does now — the demonstrated capability the ambition has to be earned from.

What a miss costs
$32,288Estimated average cost per repurchase demand. National Mortgage News, drawing on STRATMOR Group analysis.
What the rate hides
1.50%Critical defect rate, CY2025, flat against 1.52% — while eligibility defects rose 291% and credit 166%. ACES Mortgage QC Industry Trends.
What a loan costs to produce
$11,109Fully loaded total production cost per loan — not QC cost. MBA Quarterly Performance Report, Q3 2025. No per-loan QC figure is published industry-wide.
What changes
Every loan, not a sampleFull-file, cross-document checks on all production, so coverage stops being the variable that quietly gives.

How E3 works

Seven stages, and two boundaries

The model reads. Versioned rules decide. A person resolves the exception — and every step is kept with the finding.

How E3 works Seven stages. Intake, Classification and Extraction are performed by a vision-language model. Synthesis and Validation are deterministic and contain no model. Review is performed by a person. Evidence is the retained record. MODEL PROPOSES · PERCEPTION 01Intake Loan file receivedand queued. 02Classification Documents identifiedand segmented. 03Extraction Values located, keptwith page and source. 04Synthesis Reconciled across thefile; conflicts recorded. RULES DISPOSE · NO MODEL 05Validation Versioned rules run.Skips recorded too. 06Review Exception opened androuted to a named person. 07Evidence Finding kept with therule version that made it. PEOPLE DECIDE · RECORDED A model reads. Versioned rules decide. A person resolves the exception — and every step is kept with the finding.
Reading, reconciling, deciding, resolving — and what is kept from each stage.

What E3 is

Most defects live between the documents

Experienced reviewers already do this work, and do it well. Income appears on a W-2, a paystub, the 1003 and a 1040, and a good QC analyst knows to reconcile them. The constraint is not skill — it is that the checks are numerous, the file is long, and the time is finite.

So the trade gets made quietly. Random sampling is designed to estimate a rate across a population, and it does that well; whether it surfaces a defect concentrated in one channel, product or underwriter depends on how large that pocket is and how the sample was designed. Guidelines move faster than checklists. And when volume rises faster than headcount, coverage is the variable that gives, because it is the only one with no accounting entry.

E3 runs those same checks on every loan rather than a sample, and writes down what it did — the document, the page, the position on that page, and the numbered version of the rule applied. It does not replace the reviewer's judgement; it removes the reason to ration it.

The question you now have to answer

“You are using AI. Show us what it did, and how you knew it was right.”

Fannie Mae's Lender Letter LL-2026-04 took effect on 6 August 2026. Freddie Mac's Bulletin 2025-16 has been live since 3 March 2026. Both require a documented AI/ML governance programme — and both hold you to the same standard for the AI your vendors use.

For most lenders that question is uncomfortable, because the honest answer is a description rather than a record. A model produced an output, a reviewer accepted it, and what survives is a disposition with no account of how it was reached. Describing a safeguard is not the same as producing one.

E3 is built so the answer already exists. Every value carries the document, the page and the position it was read from. Every finding carries the numbered version of the rule that produced it. The rule applied and its version are fixed and do not vary between runs — so given the same inputs, the same result follows, and any disagreement can be traced to a specific value on a specific page rather than to the system as a whole.

That record is not a report written afterwards to satisfy an examiner. It is a by-product of doing the review at all, which is the kind of evidence that tends to survive being questioned two years later.

FIGURE The questions you now have to answer Fannie Mae LL-2026-04, effective 6 August 2026. Freddie Mac Bulletin 2025-16, live since 3 March 2026. WHAT THE GSEs RESERVE THE RIGHT TO ASKWHAT E3 PRODUCESWhy is AI being used?To read documents — a perception task.Named and bounded per stage.For what purpose?Locating values, not deciding outcomes.The disposition comes from a stored rule.What safeguards exist?The deciding path contains no model.Structural, not procedural.How do you know it was right?Page, position and rule version, per finding.The review can be run again.The obligation extends to vendor AI, held to the same standard as your own. E3 produces the evidence; the lender holds the obligation. No approval or endorsement is implied.
The four questions, and what E3 produces in answer to each.

E3 produces the evidence; the lender holds the obligation. Nothing here implies approval, certification or endorsement by Fannie Mae or Freddie Mac.

Controls

The policy E3 enforces, encoded and versioned

Not a checklist held in someone's head. Each control is a stored rule, bound to a loan programme, and the record shows which ran and which were skipped.

TRID
Disclosure content and timing
Loan Estimate within a period of application; Closing Disclosure received a required interval before consummation; tolerance comparisons between the two.
ATR / QM
Ability to repay
A documented, verified determination — with the points-and-fees test and the safe-harbour threshold applied by loan pricing.
HOEPA
High-cost triggers
Additional protections and restrictions where rate or fee thresholds are exceeded.
HMDA
Reported data accuracy
Loan Application Register values checked against the file they were drawn from.
RESPA
Settlement services
Section 8 referral and unearned-fee prohibitions; Section 10 escrow administration.
ECOA / Reg B
Adverse action and timing
Notice content and the period within which it must issue.
FCRA
Consumer reports
Permissible purpose and risk-based pricing notice timing.
Agency programme
FHA · VA · USDA
Programme eligibility and the layered documentation logic that stacks rather than replaces.
Selling guides
Investor requirements
The operative standard for conforming loans, encoded as versioned rules rather than remembered.

Your data

We do not train on your borrower documents

The question every lender asks second, so it belongs on the front page rather than in a contract schedule.

TRAINING
Your documents are not training data
Borrower documents and the findings drawn from them are not used to train any model — not ours, and not the model provider's. Any future domain model is trained on licensed and synthetic corpora, or on customer data only under a separate, explicit, written agreement.
PROCESSING
A third-party foundation model reads the documents
Document content is sent to a general-purpose vision-language model hosted in Amazon Bedrock, inside our AWS environment. We say this plainly because the alternative — implying a private model — would be false.
SEPARATION
Tenant-scoped access control
Access is scoped by tenant and role. We describe this as access control rather than infrastructure isolation, because that is what it is.
RETENTION
Retention and deletion are contractual
Retention periods, deletion on request and end-of-term destruction are set in the agreement, not left to a default. Ask for the current terms and we will send them before a pilot, not after.
VENDOR GOVERNANCE
Diligence answered in writing
Because your obligations under LL-2026-04 extend to us, we expect a vendor assessment and will answer it in writing rather than pointing at a web page.

Who we serve

The same file. Different reasons for caring what is in it.

Eleven segments — not eleven pages saying the same thing with the audience noun swapped.

Third-party QC firms
You run to your clients' rule sets, several at once. Your product is not the finding — it is the evidence behind it, because your client will be audited on it.
Correspondent aggregators
You carry risk on loans you did not originate, in a short window between purchase and downstream sale.
Wholesale lenders
File quality is set by brokers before the file reaches you. You inherit what was submitted and answer for it.
Independent mortgage banks
Volume swings and the team does not. Coverage is the thing that quietly gives.
Non-QM and specialty lenders
Bank statements, DSCR and asset depletion, where conventional agency rule sets simply do not apply.
State housing finance agencies
Programme rules, income and purchase-price limits sit on top of ordinary loan quality.
Government lending
FHA, VA and USDA logic that stacks rather than replaces — 203(k), manual underwrites, streamline refinances.
Community banks and credit unions
The same regulatory weight as the largest banks, with a fraction of the headcount.
Large bank mortgage divisions
QC as institutional risk control, across multiple origination channels.
Regional mortgage corporations
Distributed branches across states, where practice drift is hard to tell from jurisdiction.
Servicers and audit QC
Transferred files, escrow and payment-history validation, and post-sale defect identification.

Bring a recent QC file.

The useful first conversation is a working session on your own documents. You see how each value was read, which rule was applied, and what the record looks like when someone asks in eighteen months.

Schedule a demo Contact us