-
-
Harness scans the listing for claims made about property & reasoning on how to verify this data & pulling required data from CoStar sources
-
Extension Side panel showing the final verdict and listing mismatches with confidence score & artefact citations.
-
Full RealtyCheck Report - 1: Landing Page
-
Listed v/s Reality - 1: Showing the correct fact and what was lied about.
-
Listed v/s Reality - 2: Showing the correct fact with 3d model annotation & busting the wrong one.
-
Listed v/s Reality - 3: Model annotations in full 3d walkthrough.
-
Shows what the listing left out and drafts questions to ask before you commit.
-
Claim-wise artifact comparison views
-
Artifact Library - The artefacts that harness pulls-in for verifying all claims.
-
System Architecture
Inspiration
We love to travel, and we book our stays the way everyone does. We look at twelve photos, read a paragraph written by the person who wants our money, and pay.
You know how that goes. You open the door and the room is smaller than it looked. The bedroom is not as big as the photos made it seem. It is the same when you rent an apartment, buy a house or lease an office. The person selling writes the listing, and nobody checks it.
The truth is already on file with CoStar, Matterport, Vexcel and county records. We built Realty Check to bring it to the listing, so you get a reality check before you pay.
What it does
Realty Check is a custom agent harness inside a Chrome extension that fact-checks a property listing. Click the address on any listing page. About a minute later, a drawer shows each claim that does not match, with the value on record, a confidence score and the source that decided it.
The full report opens the 3D scan at the exact spot that proves the point, shows what the listing left out, and drafts questions to ask before you commit.
How we built it
The harness is the product. We built it on LangChain deepagents and changed almost every default.
- One orchestrator, three subagents in parallel.
insideworks from the Matterport scan,outsidefrom Vexcel aerial data, andpaperfrom CoStar and county records. - We took tools away. No file, shell or to-do tools. The agents can call our own tools and nothing else.
- Our own model adapter. Claude runs headless behind a LangChain chat model we wrote. The project holds no API key.
- Hard limits. 14 model calls and 96 seconds per run.
- The model never judges. It gathers evidence, and rule code decides the verdict.
- Every verdict needs proof. No citation means "couldn't check".
Challenges we ran into
False alarms. Our first version flagged 20 mismatches across ten real listing sites, and most were wrong. After we rewrote the rules, the same ten sites produced 2.
Speed. One check took 147 seconds. Parallel subagents and caching brought it down to 56.
Accomplishments that we're proud of
- All four sources work together in one run. Floor counts are settled by roof height measured from the air, and unit sizes by the area measured inside the 3D scan.
- The 3D viewer opens at the exact spot that proves a claim, not at the start of the tour.
- It works on any listing site, with no site-specific code.
What we learned
An agent is good at finding evidence and poor at being the judge. Most of a harness is limits.
What's next for Realty Check
We tested on apartments in Houston and Los Angeles. Next come short stays, homes for sale and commercial space, then an API so any platform can check a listing before it is published.
Built With
- chrome
- costar
- custom-harness
- deepagents
- langchain
- matterport
- playwright
- server-sent-events
- vexcel