Disclaimer: Independent product concept by Kaushal Khodifad. Not a live commercial product.return to portfolio
Disclaimer: Independent product concept by Kaushal Khodifad.
The working product

US Residential Mortgage

The live prototype first - the market and the spec are below.

Arbiter UnderwriterLive prototype ↓
Live prototype · multi-agent

How the agents work together

One file, the whole fleet - a safeguard on every handoff.

Watch one borrower file flow through the fleet
cited evidence objects
cited decision + conditions
Guardrail
PII redaction + tamper check
Measured
extraction F1 99.2%
Human checkpoint
None - runs autonomously here

Reads every document, extracts fields with provenance, flags tampering.

Always on · across every stage
Policy-as-code governs every stepImmutable, replayable audit trailConfidence signal gates every decisionHuman-in-the-loop below the floorHallucination measured · grounded-or-abstainEvery step recoverable / rollback-safe
Live prototype · agent console

Watch a single agent in depth

Fully instrumented. Run it live.

Arbiter UnderwriterL2Residential mortgageSynthetic
Illustrative · synthetic - not measured results
58%
Auto-cleared
of clean conforming files
1.8s
Draft latency
p50 / 6.2s p95
100%
Grounded
cite-or-abstain
0.18%
Income defect
vs ~1.1% baseline
Conforming purchase · self-employed borrower · A. Rivera - sole proprietor (Schedule C)
  1. 1INTAKE
  2. 2EXTRACT
  3. 3DERIVE
  4. 4TEST
  5. 5CHECKPOINT
  6. 6LEDGER

Classify the file & documents

Auto

Identified a conforming purchase with self-employed income; loaded 18 documents.

tool callclassify_file()
in: 18 docs (URLA, 1040s, P&L, AUS)
out: {loanType: conforming-purchase, employmentType: self-employed}
PII redaction - SSN + account numbers masked
Confidence signal99%
INTAKE-OK
How it works

The market, from zero

No finance background needed - tap any step.

01Application (Apply)

Day 0; often completed in 1-3 days

The borrower fills out a standard form (the URLA / Form 1003) telling the lender who they are, how much they earn, what they own, and which house they want to buy. Think of it as the master intake form for the whole loan.

Who: Borrower + Loan Officer (the salesperson at the lender who guides them)

Where we play

Where Arbiter fits

The first workflow to automate - and where a human stays in.

The thin-slice wedgeArbiter Underwriter

Automated income calculation on conforming purchase loans

#1
income = top defect category
~$11k
cost to originate per loan
repurchase
the risk that kills lenders

Income calc is the most error-prone step, and bad math forces costly GSE buybacks.

What the agent decides
  • Classify each income source, pick the method
  • Average variable bonus & commission income
  • Reconcile the calc vs 1003 and AUS
Where a human stays in
  • Underwriter signs the final qualifying income
  • Human adjudicates declining self-employed income
Build-ready

The PRD

Tap any section to expand it.

Arbiter Underwriter
From thin slice to governed digital worker
One high-leverage workflow, scoped tight enough for an engineer to build.
1Scope
2Build
3Govern
4Ship
The cast

Who's involved

Who touches the deal, and what each does.

Borrower
The person taking the loan to buy a home

The homebuyer. They supply all the documents, pay closing costs and a down payment, and make the monthly payment for years.

Loan Officer (LO) / Mortgage Broker
Front-line salesperson who originates the loan

The person who helps the borrower apply, picks a loan product, and quotes a rate. A loan officer works for one lender; a broker shops multiple lenders. They earn commission, so they're motivated to get loans to close.

Loan Processor
Assembles and organizes the loan file

The behind-the-scenes coordinator who collects every document, orders the appraisal and title, and chases missing items so the file is complete before a decision.

Underwriter
Makes the approve/deny decision

The risk gatekeeper. They verify the borrower can actually repay and the file follows the rulebook, then approve (often 'conditionally') or decline. The most judgment-heavy human in the chain.

Appraiser
Independent valuer of the property

A licensed, neutral expert who decides what the home is actually worth so the lender doesn't over-lend. Required by law to be independent from the loan's sales side.

Lender / Mortgage Originator
The company that funds the loan

The business whose money actually buys the house at closing. Can be a bank or an independent mortgage bank (IMB). Most lenders sell the loan soon after, recycling their cash into the next loan.

Fannie Mae & Freddie Mac (the GSEs)
Government-sponsored buyers of conforming loans

Two huge quasi-government companies that buy 'normal' (conforming) mortgages from lenders, guarantee them, and bundle them into bonds. They effectively write the rulebook (Selling Guide) that most US loans must follow. GSE = Government-Sponsored Enterprise.

Ginnie Mae
Guarantor for government-backed loan bonds

A government agency that doesn't buy loans itself but guarantees the bonds made from FHA and VA (government-insured) loans, making them safe for investors. Backed by the full faith of the US government.

MBS Investors
Buyers of the mortgage bonds

Pension funds, insurers, banks, and foreign governments who buy mortgage-backed securities for steady returns. Their money is the ultimate source of nearly all US mortgage funding.

Mortgage Servicer
Collects payments and manages the loan day-to-day

The company the borrower actually pays each month. It distributes the money, runs the escrow account for taxes and insurance, and handles hardship cases. Often a different company than the lender or the loan owner.

Title / Escrow Company
Neutral closing agent

A neutral third party that confirms the seller really owns the home (clear title), holds everyone's money safely during the deal, and handles the signing and transfer of ownership at closing.

The paperwork

Every document, decoded

Tap a document - what it is, why it matters, an example.

URLA / Form 1003 (Uniform Residential Loan Application)

The master application form every borrower fills out. It captures who you are, your job and income, what you own and owe, and details of the home and loan you want. Standardized by Fannie Mae and Freddie Mac so every lender uses the same form.

Why it matters

It's the single source of truth the entire loan is built on. Every other document exists to prove what the borrower wrote here is true.

Example

Jane Smith, SSN 123-45-6789, born 1988; employed as a nurse at City Hospital for 4 years earning $7,500/month base plus $500/month overtime; has $40,000 in a Chase savings account; applying for a $360,000 loan to buy 12 Oak Street for $450,000.

The money

What it costs

Who pays for what - illustrative figures.

Borrower closing costs
Borrower
~2% to 5% of the loan amount (often ~$5,000-$7,000 on a typical single-family home)

One-time fees paid at closing: lender origination fee, appraisal, title insurance, escrow setup, recording, and prepaid taxes/insurance. On a $360,000 loan, 2-5% is roughly $7,200-$18,000.

Down payment
Borrower
Typically 3% to 20% of the home price

The borrower's own cash put toward the purchase, separate from closing costs. As low as 3% on some conforming loans; under 20% usually triggers extra mortgage insurance.

Lender cost-to-originate (cost to produce one loan)
Lender
~$10,000-$11,000 per loan (MBA: $11,076 total production expense per loan in 2024)

What it costs the lender to make one loan: loan officer commissions, processing/underwriting labor, technology, compliance, and overhead. This is the cost an AI digital worker most directly attacks.

Lender net production profit (or loss) per loan
Lender
+$443 per loan in 2024 (vs. a loss of $1,056 per loan in 2023, per MBA)

Margins are razor-thin and swing between profit and loss year to year, which is exactly why lenders are desperate to cut the ~$11k cost-to-originate.

Ongoing monthly payment (PITI)
Borrower
Varies; principal + interest + taxes + insurance, paid monthly for 15-30 years

PITI = Principal, Interest, Taxes, Insurance. The servicer collects this and routes the tax and insurance portions through the escrow account.

Scoreboard

Hero metrics

The numbers the worker has to move.

~$11,076 per loan in 2024 (MBA); ranged ~$10,700-$12,600 across recent quarters
Cost per loan (cost-to-originate)

The total it costs a lender to produce one mortgage, from labor to technology to compliance. The headline efficiency number for the industry.

~37-42 days for a purchase loan in recent reports; as fast as ~36.8 days in early-2026 ICE data
Cycle time / days-to-close

How many days from application to the loan funding. Faster is better for the borrower's experience and the lender's cost.

Roughly 76-77% of purchase applications close (ICE); pull-through has been declining over the past four years
Pull-through rate (a.k.a. straight-through / closing rate)

Of all the applications a lender takes, the share that actually make it to a funded loan. Low pull-through means wasted work and cost on loans that never close.

Income calculation (especially self-employed and rental income) is consistently the top defect category in Fannie Mae's QC samples
Defect rate

The share of closed loans found to contain a mistake (most often a miscalculated income figure) during quality-control review. Defects can force the lender to fix or buy the loan back.

Low in absolute percentage but high-dollar impact; income-calculation defects are a leading driver, which is why Fannie built a free Income Calculator tool to reduce them
Repurchase rate

How often Fannie Mae or Freddie Mac forces the lender to buy back a loan it already sold, because the loan broke the rules (often a bad income calculation). Each repurchase is very expensive for the lender.

The field

Who else is here

7 players already serving this vertical, and the gap each leaves.

ICE Mortgage Technologymoat 5/5
System-of-Record / LOS

Operates Encompass, the dominant US mortgage LOS / system-of-record, plus a servicing platform. Layering "Aurora" agentic AI and Mortgage Analyzers …

Gap: Explicitly keeps AI out of final approval/pricing/disclosure decisions, so it assists rather than decides; legacy arch…

Rocketmoat 4/5
Lender-native / Captive AI

Rocket's proprietary lender-native AI platform is Rocket Logic (and Navigator), built on 10+ petabytes of proprietary data and 50M+ annual call tran…

Gap: Primarily a captive tool optimizing Rocket's own pipeline, not an open platform competitors can buy; underwriter overs…

Blendmoat 3/5
Point-of-Sale / POS

Leading digital origination / POS platform for banks, credit unions and mortgage lenders (powered $1.2T in applications in 2024). Launched "Intellig…

Gap: Historically a front-end (conversion) layer, not the decision engine; profitability pressure as a public company; agen…

Tavantmoat 3/5
Point-of-Sale / POS

Mortgage technology and services provider behind FinXperience (collaboration POS), FinConnect (130+ integrations/data services), and the newer TOUCH…

Gap: Overlay/services model means less ownership of the system of record or final decision than ICE or in-house lender AI; …

Candor Technologymoat 4/5
Underwriting / Decisioning-AI

Loan Engineering System (LES) powered by CogniTech expert-systems / "autonomous machine" technology (US patent awarded Sept 2022). Performs income, …

Gap: Single-vertical (mortgage) point engine dependent on integrating into others' LOS/POS; smaller scale and distribution …

Ocrolusmoat 3/5
Document-AI / Extraction & Fraud

Document-AI platform that extracts and structures data from 100s of document types (bank statements, pay stubs, tax forms) at 99%+ accuracy with hum…

Gap: Sits below the decision - it verifies and feeds data but does not make the credit decision, so it is a component vulne…

nCinomoat 4/5
System-of-Record / LOS

Cloud bank operating system (on Salesforce) spanning onboarding, account opening, loan origination, credit analysis, and portfolio management. nIQ D…

Gap: AI is largely an assistive copilot/data-recognition layer, not autonomous decisioning - bankers still drive credit dec…

Cheat sheet

Jargon, decoded

Every term on this page - search it.

Conforming loan

A mortgage that follows Fannie Mae and Freddie Mac's rules and stays under their size limit, so it can be sold to them.

e.g. A $360,000 loan that meets credit, income, and size guidelines is conforming; a $2 million loan exceeds the limit and is 'jumbo' (non-conforming).

GSE (Government-Sponsored Enterprise)

A quasi-government company chartered by Congress to support a market; in mortgages this means Fannie Mae and Freddie Mac.

e.g. When a lender says 'the GSEs require two years of tax returns,' they mean Fannie Mae and Freddie Mac's shared rulebook demands it.

Secondary market

The marketplace where lenders sell loans they've already made to bigger buyers, instead of holding them.

e.g. A local lender sells your fresh mortgage to Fannie Mae a month after closing so it gets its cash back to lend again.

MBS (Mortgage-Backed Security)

A bond made by pooling thousands of mortgages together so investors can buy a slice of the monthly payments.

e.g. A pension fund buys $10 million of a Fannie Mae MBS and earns returns from the combined monthly payments of thousands of homeowners.

Securitization

The process of bundling many loans into a tradable bond (an MBS) and selling it to investors.

e.g. Freddie Mac takes 5,000 conforming loans, guarantees them, and securitizes them into a bond it sells on Wall Street.

AUS (Automated Underwriting System)

Software that instantly checks a loan against the rules and recommends approve or deny.

e.g. The lender runs the file through Desktop Underwriter and gets back 'Approve/Eligible' in seconds.

Desktop Underwriter (DU) / Loan Product Advisor (LPA)

The two main automated underwriting systems: DU is Fannie Mae's, LPA is Freddie Mac's.

e.g. A self-employed borrower's file is run through LPA because Freddie Mac's system is historically more flexible on non-traditional income.

Underwriting

The process of evaluating whether a borrower and property are safe enough to lend on, ending in an approve/deny decision.

e.g. During underwriting, the underwriter recalculates the borrower's income from tax returns and confirms the appraised value covers the loan.

Escrow account

A holding account the servicer uses to collect a slice of each monthly payment and pay the borrower's property taxes and insurance when due.

e.g. Jane's $2,400 monthly payment includes $400 that goes into escrow, which the servicer uses to pay her $3,600 property tax bill and $1,200 insurance bill each year.

PITI

The four parts of a typical monthly mortgage payment: Principal, Interest, Taxes, and Insurance.

e.g. A $2,400 PITI payment might be $1,500 principal+interest, $700 taxes, and $200 homeowner's insurance.

Loan servicing

The ongoing job of collecting payments, running escrow, and handling the loan day-to-day after it closes.

e.g. Three months after closing, Jane gets a letter saying her servicing has transferred and she now pays a company she's never heard of.

Qualifying income

The reliable monthly income figure an underwriter calculates and uses to decide how big a loan the borrower can afford.

e.g. A freelancer reports wildly different yearly earnings, so the underwriter averages two years of tax returns to set qualifying income at $6,250/month.

Defect

A mistake found in a closed loan during quality-control review, most commonly a miscalculated income figure.

e.g. A QC review finds the underwriter counted a one-time bonus as recurring income, creating an income-calculation defect.

Repurchase (buyback)

When Fannie or Freddie forces a lender to buy back a sold loan because it violated the rules.

e.g. Because the income was overstated by a defect, Fannie Mae demands the lender repurchase the $360,000 loan at full price.

Pull-through rate

The share of loan applications that actually close and fund.

e.g. If a lender takes 100 applications and 77 close, its pull-through rate is 77%.

Conditional approval

An underwriter's 'yes, but' - approval that depends on the borrower supplying a final list of items.

e.g. The loan is approved on the condition the borrower provides one more pay stub and a letter explaining a $5,000 deposit.

LTV (Loan-to-Value)

The loan amount as a percentage of the home's value; lower means more borrower equity and less lender risk.

e.g. A $360,000 loan on a $450,000 home is an 80% LTV, meaning the borrower put 20% down.

↑ Independent product concept by Kaushal Khodifad. Arbiter is an independent product concept by Kaushal Khodifad; it is not a real company or a commercial product. It explores the adjudication infrastructure for regulated lending & insurance space. Not a live commercial product. Data is illustrative.

Return to portfolioOpen the live demo