CoStar HackPrinceton kit: build a richer per-building record by joining property, tenancy, company and sale files
Recipe that worked in about 8 s on the 410 MB kit (streaming JSONL, no pandas): 1) commercial-properties.csv maps PropertyId to MatterportModelId (several rows per property). 2) data/commercial-property.jsonl: buildingName, primaryAddress.latitude/longitude, researchMarketId (14 = Houston, 20 = Los Angeles), numStoriesAboveGrade, propertyType (enums.json PropertyType: 5 Office, 6 GeneralRetail, 11 Multifamily, 1 Hospitality...). Year built = constructionEvents[] with eventType 2 (ConstructionComplete); 4 = RenovationFinished. notes[].value holds the research prose. 3) Tenants: location-occupancy.jsonl rows with propertyId and tenantLocationId, then company-detail.jsonl id -> name. Sort by fromDate desc to get recent tenants first. 4) Sales: commercial-sale-comp.jsonl entities[].propertyId -> soldDate, soldForPrice. About 2,700 of the 3,339 properties have a name, year built and a Matterport model. Filter tenant names before showing them to users: a few real company names are profane.