CoStar: Inside and Out, AI for the Built World
All three layers in one frame (Matterport interiors, Vexcel aerial and 3D city, CoStar records), and an agent that plans, fans out per building and checks claims across layers.
Judged 30% use of the data, 30% agentic depth, 25% usefulness, 15% demo and polish.
Angle. The track asks for something that makes CoStar say “we didn’t know you could do that with our data”. Every CoStar product today answers questions inside one listing. Cranley answers a person’s question across a whole city, by measurement, with all three layers checking each other.
| Criterion | What serves it |
|---|---|
| Use of the data (30%) | All three layers in one frame: walls and doorways from Matterport, elevation, ortho, property facts and the 3D city from Vexcel, records from CoStar. Cross-layer tools: record versus reality (listed vs measured area, claims check), flood (Vexcel ground elevation + Matterport door sills), garage and driveway fit (interior garage + aerial driveway), site tools (docks, truck access, parking from the air), drone splat aligned to the scan. Search by measurement over every scan with a wall graph. |
| Agentic depth (30%) | A deepagents investigator that plans (todo list the person watches), fans out one subagent per building in parallel, asks before expensive steps, remembers what it learned per building, and resumes after a restart from a MongoDB checkpoint. The composer picks tools for the need from more than 75, asks only the missing inputs, and the workspace is a document the agent and the person edit with the same operations (with undo). |
| Usefulness (25%) | 50 personas, each a real user with a named need: wheelchair user, fire captain, Houston homeowner before a storm, truck dispatcher, office tenant, CoStar researcher, insurance underwriter, daycare operator, car owner. For the people who open CoStar every morning: a tenant rep’s weighted site scorecard across candidates, market comps from CoStar records, industrial fit (clear height, docks, truck court), retail site screening, commute fit for a whole team, and a facility audit PDF. PDF report a person can hand to a landlord. |
| Demo and polish (15%) | Photoreal building on its site, flood rising in the street, smoke filling rooms, the agent’s lanes working in parallel. A dense working tool, not a showcase. |
Scripted flow: web/src/skins/app/demo/flows/costar.json, “Check a listing against the scan, then tour it”: the agent reads the
listing, the CoStar record and the scan side by side, the camera goes to the room where they disagree, then the tour
plays every finding. The wheelchair path below is the spoken alternative.
Demo path (2 minutes).
- Home. Type “I use a wheelchair, 81 cm wide. Find a one-bedroom in Houston I can get into and use the bathroom.”
- The composer adds step-free route, door widths, turning space and bathroom clearance, asks nothing (width given).
- The investigator’s lanes run five buildings at once; the ranked table fills.
- Open the top building: photoreal, on its site in the city. The route draws from the curb to the toilet; the one narrow door shows red with “Matterport geometry, high”.
- Record versus reality on the same building: the listed area against the measured one.
- Twenty-second switch: Houston homeowner, Harvey water rising through the same photoreal house, the list of what to move first.
- Close on the catalogue and the line “every number from a calculation”.