The property tax a rental house will owe after you buy it, by address

Each house bought by a private single-family-rental operator with 30 to 300 homes is underwritten in Excel, one line at a time. Its tax line is usually the seller's bill, wrong within months of closing wherever the county reassesses after a sale.

Bearing Tree is a lookup platform, web app and API, for an underwriting model's property-tax line. From an address, a purchase price and a closing date, it finds the parcel, reads its latest bill, applies the county's rule row and returns a dated Excel row: the seller's bill, the reassessed bill as a low and a high, and each figure's source.

Bearing Tree needs a model to tie typed addresses to parcels and to read county bills in whatever layout the county serves. About one address in five so far escapes exact matching, new construction on a parent parcel or a condominium unit among them; an embedding model over address, legal description and neighboring parcels ranks candidates. A fine-tuned document model reads each levy from HTML, scanned PDF and image-only bills.

At Bearing Tree, people settle every parcel match below 0.92, every bill that does not add up and every county's rule row. Each parcel or bill decision leaves a labeled pair, a typed address against its parcel number or a bill region against its confirmed value, with no customer name, price or date. Match pairs train the next matcher and bill pairs the next reader, so fewer rows wait; the corpus is ours.

A new county, a bill season or a customer's pipeline list is what keeps Bearing Tree's GPUs busy. A county is several hundred thousand documents parsed in a few days, a bill season re-reads whole counties within weeks, and a 2,000-row list is due back before the bid deadline. Between them the GPUs wait. A purchase price tied to an address is a live bid, so submissions stay with our models on cloud infrastructure in Ohio.

The parcel matcher and the bill reader are Bearing Tree's own fine-tunes of open-weight models, run by the hour on cloud GPU capacity we control in Ohio. By December 2026 an embedding-retrieval index over every covered parcel replaces the current matcher, so a list matches in one pass. Rule-row drafting, reading only public sources, leaves another provider's hosted model for an open-weight model on the same capacity in June 2027. Past 150 counties, reserved capacity replaces hourly rental.

Bearing Tree Data, Inc. is a Delaware corporation founded in October 2024, with its team in Cincinnati and its service live. The rule table has taken most of the time since: no county's row is published until an editor signs it, which is why coverage opened at 38 counties.