LevelXero is the transaction rail for residential real estate. Whether you are taking a listing or writing an offer on one, the deal runs on a single structured record: priced from public county data, papered with the right forms for the jurisdiction, signed in product, and carried through to close. The listing agent, the buyer's agent, both clients, and the lender all work from that one record instead of a pile of PDFs.
By the founder Nicole Beaulieu, PhD, Computer Science · Formerly CTO at Figure Technologies · Inventor on 45+ issued US patents
We ran the engine against a real Reno home that sold in March for $1,110,000, with the property excluded from its own comparable set. It drew twelve verified arms-length sales from the county record and returned a range whose top landed within 0.6% of the actual sale price. No MLS, no valuation vendor, no model anywhere in the number.
A deal can start four ways: an agent takes a listing, a seller comes in directly, an agent writes an offer, or a buyer comes in directly. All four land on the same record. When the seller and the buyer are both here, the offer is structured data on both ends rather than a PDF in an email thread, and the negotiation stops being a thread at all.
LevelXero works for the licensed professional, not around them. AI drafts, explains, and watches the clock; the system commits, and a human decides everything that binds. There is no prompt to learn. The surfaces below are marked with what is actually built today.
The address resolves against our own parcel store, so the facts come from the county rather than from memory. Those facts flow into the listing, the disclosures, the agreement, and the MLS sheet without being retyped. Repeat entry is the loudest complaint agents make about their tools, and here it is simply gone.
Comps come from the county's recorded sales, filtered by the county's own validity codes, so a family transfer or a foreclosure is excluded by name rather than by guesswork. The agent adds and removes comps; the range recomputes. What to list at and what to offer are the same question run in opposite directions, against the same evidence.
Say or type what is good about the house. The AI returns structured MLS fields, public and private remarks, and showing instructions as a proposal with its reasoning. The agent edits and confirms. If the model is unavailable the form still works, by hand.
Nevada's disclosure set and the listing agreement are filled from the listing's own facts, using versioned form definitions that a human reviewed once. No model touches a legal field. A new state is a data change, which is why this reaches beyond Nevada.
The packet is signed where it was created, and publication runs as a durable job: queued, retried with backoff, and held until the board accepts it. It does not hand the agent a copy-paste chore when something upstream is slow. Re-keying is the tax this product exists to remove, so a publication that needs a human needs us, not the agent.
Inspection, financing, appraisal, title: every contingency tracked from execution, every inbound document read and checked against the record, every follow-up drafted. This is the work agents pay $350 to $450 a file to hand off, and it is what the structured record exists to make possible.
Typeahead resolves against our own parcel store rather than a geocoder, so a selected suggestion always has a real assessor record behind it. Beds, baths, area, lot, vintage, and the assessor's own market-area code come back typed and ready to fill a form.
In a county we have not loaded yet, the agent types the facts in and nothing is blocked. Every outside dependency here falls back to doing it by hand rather than stopping the deal.
{
"jurisdiction": "NV-washoe",
"beds": 4,
"baths": 3.5,
"building_sqft": 2914,
"lot_size_sqft": 7667,
"year_built": 2003,
"stories": "2 Story",
"grade": "Very Good",
"neighborhood": "RBFD", // assessor market area
"assessed_value_cents": 23702600,
"legal_description": "…",
"source_run_id": "…" // which nightly ingest wrote it
}Rough notes go in, typed or spoken. What comes back is not prose we hope is right: it is a schema-validated proposal in which every field declares which of your inputs informed it. The schema forbids extra properties, so the model cannot invent a fee, a name, or a number you did not give it.
Private details stay private by construction. The agent edits and confirms; only then does the deterministic system commit. If the model is unavailable, the form still works by hand.
// in: whatever the agent says about the house
"remodeled kitchen, quartz, mountain view,
lockbox on the gas meter, dog on site"
// out: every field cites its grounding
{
"public_remarks": {
"text": "…",
"grounded_on": ["rough_features", "parcel_facts"]
},
"private_remarks": {
"text": "Lockbox on the gas meter. Dog on site.",
"grounded_on": ["rough_features"]
},
"amenities": { "items": […], "grounded_on": […] }
}
// additionalProperties: false, so no invented fieldsNevada's disclosure set lives in the database as versioned form definitions, and every field declares its own source. Filling a form resolves those definitions against the listing's facts using a fixed, reviewed set of conversions rather than free-form code. No model touches a legal field, and every filled document records the version of the definition that produced it.
Conditional forms are data too: answering yes to question 2g on the seller disclosure triggers the open-range disclosure automatically. A new state is a reviewed data change, which is the whole reason the product travels beyond Nevada.
524 Consent to Act
525 Duties Owed by a Nevada Licensee
525a Duties Owed, supplemental
547 Seller Real Property Disclosure
551 Open Range Disclosure
↳ triggered when 547 Q2g = yes
610 Manufactured Home Disclosure
↳ triggered by land use
ERTS Exclusive Right to Sell
↳ the agreement activation binds
// a new state is a data change, not a releaseOnce the transaction record carries the deal, the financing handoff is a consequence rather than a project. On contract execution LevelXero emits a MISMO 3.4 payload to the lender's webhook: property, parties, financing terms, contingency clocks, and the audit chain back to the listing. All typed, all validated, all before the borrower's phone buzzes.
Every lender's loan origination system has waited two decades for exactly this, and it is why the record underneath the product is built the way it is.
{
"event": "contract.executed",
"mismoVersion": "3.4",
"transactionId": "a4f8c2e1-9b2d-4f1a-8c3e-d72f1a8b9d3f",
"property": {
"address": "117 Bristlecone Ct, Reno, NV 89519",
"apn": "043-271-15",
"jurisdiction": "NV"
},
"purchasePrice": 127500000, // cents
"closingDate": "2026-06-30",
"financing": {
"lenderId": "bay-cap-mtg",
"loanAmount": 102000000,
"downPayment": 25500000,
"contingencyDeadline": "2026-05-29"
},
"parties": {
"sellerOfRecord": "prin_01HE...AB12",
"buyerOfRecord": "prin_01HE...CD34"
},
"auditChain": "sha256:9f3a...2c4e"
}For the first time, the full transaction record exists as structured data. Not just the outcome. The arc, the timing, the terms that moved, the terms that held. A dataset no one else can build, because no one else sits where the deal is created.
The product is the surface. The data is the moat.
Nicole Beaulieu. Six years at Figure Technologies, architecting the embedded lending platform behind 100+ private-label partnerships. The platform now drives $1B+ in monthly loan volume. I joined as a Principal Engineer in 2019 and worked up through Director, VP Engineering, and CTO. I built the engine before I ran it.
Inventor on 45+ issued US patents in gaming and platform systems. Principal inventor on a blockchain-based mortgage assignment registry (DART, at Figure). PhD, Computer Science (UNR, 2025); the dissertation formalized 30+ years of building the infrastructure on which other companies' products run.
The lender side I've already built. The origination side is what I'm building now.