Rivl Real Estate is a Missouri real estate brokerage that we hold the license for and that we operate. A seller enters an address. AI drafts the listing, public county records set a price range, a licensed broker reviews every listing before it publishes, and a buyer makes an offer through a constrained composer that cannot produce a term nobody approved.
We did not build it to sell a real estate product. We built it to prove that a regulated workflow can be run end to end by a system that is accountable for what it says.
The seller sees the listing exactly as a buyer will see it before it goes anywhere. Every edit they make is screened again by the same compliance gate that screened the original.
The interesting step is the third one. It is the only one that exists to say no, and it is the reason the other five can be automated at all.
The seller signs in through a single use link and enters the property. No account password to lose.
The model writes the listing from the facts on file. Nothing it cannot trace to an entered fact is allowed into the copy.
Two layers, a pattern detector and a model judge, screen for fair housing risk. A hard flag blocks. A soft flag routes to a human.
An engine builds a range from arm's length county sales, adjusted for living area and recency, with the comparable count shown.
A licensed broker reviews and publishes as a separate deliberate act. Approvals are born pending and cannot be forged.
Buyers compose offers from approved terms only. Counter, accept and reject run through the same record.
Rivl Real Estate operates as a transaction broker, a single neutral intermediary rather than an agent for either side. The recommended range is an approximation built from public records and is not an appraisal or a valuation opinion.
Anyone can put a model behind a form. What separates a demonstration from a system that a regulator, a lawyer and a customer can all live with is the machinery underneath that makes the wrong outcome impossible rather than unlikely.
Generated copy runs against a fixed corpus of scored cases on three separate axes, with deliberately planted negative controls that must be caught. We found that one axis had been measured with only half the production screening in place, corrected it, and watched the score fall below the bar. The release stopped. That is the gate working.
An approval cannot be created already approved. A listing cannot publish without a stored executed agreement. A notice cannot claim delivery that never happened. These are constraints in the schema, so a bug in the application, or a party calling the interface directly, still cannot produce the forbidden state.
Every material act, generation, screening, edit, approval, offer, counter and acceptance, is recorded and exportable as one document. When a correction is made after the fact it is labeled as a correction rather than quietly overwritten.
The pricing engine loads county assessor sale records, uses the county's own arm's length classification instead of one we invented, separates the file date from the data vintage so a load cannot contain sales newer than the file that carried them, and refuses to publish a range without a minimum comparable count.
The offer composer draws from a whitelist held as data, with a database relationship that makes an unapproved blank unrepresentable. There is no free text cover letter and no path from the composer to a model. The buyer is derived from the session rather than passed as a parameter, so one buyer cannot compose as another.
Notification status is stored per message with the outcome the email provider actually returned. The system reports what happened rather than what it attempted, including when the answer is that delivery failed.
Real estate advertising is regulated speech with a licensed accountable human at the end of it. So is a claim denial letter, a patient financial notice, a credit adverse action, and a suitability disclosure. The machinery is the same machinery.
Any place your organization produces customer facing language that a regulator could read back to you. Denial and appeal letters, adverse action notices, disclosures, plan summaries. Screened before send, with a documented pass rate and a stop condition.
Approval chains, dual control, licensed sign off, segregation of duties. Enforced at the data layer so the control is real rather than procedural, and so an audit finding is a query rather than an investigation.
Examinations, discovery, internal audit and vendor due diligence all ask the same question, show me what happened and when. The answer should be a document the system produces, not a project.
Extracting, bounding and refreshing a public data source, then being explicit about coverage gaps rather than filling them with a plausible number. Most valuation, benchmarking and rate comparison problems are this problem.
Where a model must help a user act but must not be allowed to author the terms. Applications, submissions, orders, authorizations. The user gets speed. The organization keeps the boundary.
A working regulated platform in months rather than years, built the way we would build yours, with the evidence to show a compliance officer why the timeline did not cost you control.
Fifteen minutes. Bring the workflow that has to be right every time, and we will tell you which parts of this apply and which parts do not.