Automated underwriting, in production, at national scale.
Dwelling Blocks is the system behind the largest direct-to-appraiser appraisal operation in the United States. More than a thousand transactions a day move through one event-sourced order record: the order created from the loan, assignment across 15,000 vendors, borrower payment and attestation, inspection scheduling and tracking, a model that reads the report and every photograph against an unlimited rule set, exception handling, vendor payout, and delivery to both GSEs over a UAD 3.6 Direct Integration. Review is one step of many. Most of them happen without a person touching them. We are hiring the engineers who will build what comes next.
Every step of the workflow is a place where software does work a person used to do.
An appraisal is not one task. It is an order placed from a loan file, a vendor found and vetted, a borrower who has to pay and confirm their information, an inspection that has to be scheduled and completed, a report that has to be read, exceptions that have to be resolved, a vendor who has to be paid and a file that has to reach the GSE and come back accepted. Every one of those used to be a person with a phone and an inbox. On Dwelling Blocks every one of them is structured data on one order record, and most of them are handled by a rule or a model with a reason logged.
Some of that is classic systems engineering: an append-only order record, a ledger that has settled hundreds of millions of dollars, a Direct Integration that delivered UAD 3.6 to both GSEs before most of the industry had read the spec. Some of it is applied AI: assignment that ranks 15,000 vendors by credential, coverage and performance; borrower attestation that reads what the borrower submitted and validates it against the order; an exception engine that resolves most of what used to be a phone call; and Ava, a model that reviews the entire report and every photograph against an unlimited rule set in seconds, as its own document beside the appraisal, for the underwriter and the appraiser alike.
The result is the largest structured collateral dataset any single platform sees, with what happened next attached to every order. Very few teams have shipped models into a regulated, money-moving financial workflow at national scale with evaluation data to prove they are right. We have. What comes next is built on that.
The problems
What the engineering actually is, across the whole workflow. Each of these runs in production today and each has an obvious next version.
Run a thousand orders a day through eleven steps with no one watching.
Create, assign, collect, schedule, inspect, upload, review, resolve, pay, deliver, accept. Each step is an event with a rule or a model behind it, retries and timeouts that are safe under failure, and a person who can step in at any point without breaking the chain.
"Three high-quality photos of the kitchen" has to run correctly in Texas and never in California.
Underwriters write rules the way they would say them. The system scopes, versions and applies them exactly, and generates the message the appraiser sees when one fails.
Read the whole appraisal, every form, every photograph.
Ava reviews the report, addenda, sketch and every photograph the appraiser attached, on the redesigned URAR and every legacy form, with no layout described in advance. A finding on every rule, with the evidence that produced it, in seconds.
Read what the borrower submitted and validate it against the order.
Identity, occupancy, property details and payment, collected from the borrower and checked against the loan and the appraisal before they become an exception downstream.
Let software resolve the exception, and prove afterward that it should have.
Assignment, fee bumps, missing contacts, resubmissions: handled by rules and models inside an event log with a timestamp and a reason on every action, replayable for any loan on any day.
A ledger that has settled hundreds of millions of dollars and balances to the cent.
Borrower collection at order, vendor payout on delivery through Stripe, a receipt on every transfer, idempotent under retry and reconcilable by anyone with read access.
Train the next models on the largest structured collateral dataset any single platform sees.
Orders, vendors, inspections, reports, photographs, revisions, review findings, payments and outcomes, with what happened next attached. On November 2 the appraisal becomes structured data for the first time; we have been reading it that way since 2025.
The constraints are the interesting part.
A wrong answer anywhere in the workflow is a loan that does not close, a vendor who is not paid, a borrower charged twice or a file a regulator cannot reconstruct. That shapes everything: how models ship, how state is stored, how money moves, how a person overrides the system and how that override is recorded. These are the rules the codebase is built around. They are not aspirations; they are what the tests check.
- Evaluation gates release
- Every model change runs against real reports with known outcomes before it touches a pipeline. If the numbers are not better, it does not ship.
- Evidence or it is a bug
- A finding without the passage, the photo or the field that produced it does not exist. Reviewers, auditors and examiners read the why.
- Inconclusive is a valid output
- The model is measured on knowing when it cannot read something and handing it to a person, not on always answering.
- Pinned and replayable
- Rule versions and model versions are recorded on every finding. Any decision from any day can be reproduced.
- Append-only order record
- State is an event stream. Nothing is edited in place; corrections are new events with an actor and a reason.
- Human override, recorded
- A person can override anything the system decides, and the override becomes part of the record with the same rigor.
- Separate from the appraisal
- The review is its own document and dataset. Nothing the system does modifies the appraisal or the opinion of value.
Who you would join
The founders built the platform that powers the largest direct-to-appraiser appraisal operation in the United States, moved the largest lender in the country from AMCs onto it, and shipped UAD 3.6 to the GSEs before most of the industry had read the spec. One of them is a third-generation appraiser who has written the reports the model now reads. The engineers here each own a whole piece of a system that settles money, reads evidence and answers to regulators, and every one of them can explain what the model did on any order from any day.
We hire for that standard: people who know the domain and the models better than anyone in the category, and who want to be responsible for a whole thing. About the company →
How we work
Remote, and built for it. Decisions are written down. Missing a call never means missing the decision. Hours bend around lives and time zones.
Few meetings, and real ones. A meeting needs a reason and an outcome. Everything else is a message.
Ship small, ship often. We would rather learn something true in a week than defend a plan for a quarter. A public changelog records every change, dated.
Own it end to end. You take a rough idea through design, build and the part most places skip: running it in production and watching what it does to real numbers.
Learn the domain. Appraisals, inspections, QC, GSE delivery. You do not need to arrive knowing any of it. You do need to find it interesting, because the leverage is in the details.
Say the true thing. Lenders put collateral decisions in our hands and we hold SOC 2 Type II because that trust has to be provable. Inside the team the same standard applies: raise the broken thing early.
Open roles
Software Engineer
You are an expert in large systems and in putting models to work inside a real workflow. You have run production infrastructure that cannot lose a transaction, and you know what it takes to ship a model whose output has to be right, explained and reproducible. You work on your own, decide what good looks like, and ship the best version of it without being asked twice.
Send a short note and a link to something you built that runs in production.
careers@dwellingblocks.comEnterprise Sales
You have sold enterprise software into regulated companies and you know how a lender decides. You run the process from the first call to launch: operations, underwriting, compliance and IT, each with their own questions, each needing a straight answer. You will have the most advanced platform in the category behind you, the largest direct-to-appraiser operation in the country as the reference, and a dedicated implementation team that makes every promise you make true.
careers@dwellingblocks.com