ICHOR · OS
ICHOR / Domains / Real estate & development
D-01 · Real estate & development Live
Groma

Find the land before the market does.

Groma is the ICHOR engine for real estate and development. It runs both ends of the problem as one system, sourcing and underwriting from the authoritative record itself, and it is honest about what it can and cannot predict.

The decision problem

A developer or acquirer lives on two questions: what is worth pursuing, and what is it actually worth. Answering either well means holding the whole county in view at once, then underwriting a specific parcel to institutional standard on a deal timeline.

The information is public, but it is scattered and un-indexed. A loan's maturity date lives in a recorded document image, not a database. The next rezone is buried in a council agenda. Assessor values, ownership, zoning, flood, soils, and grid capacity each sit in a different system. The deal moves faster than the stack, so the sharpest operators still miss land they should have owned and still get surprised in diligence by a problem the record already held.

You don't lack tools. You have thirty of them, and none of them talk to each other.

Watch it run

A coverage surface: developable land across a market, every parcel carrying its own record. Ownership and value here are drawn from the public parcel record, so nothing is hidden.

Groma · coverage surface · public parcel recordrecorded capture · 2026
The parcel record is public, so it runs at full fidelity. Client identity and any private book overlay are never shown.

The method, applied to land

The same five-step method behind every ICHOR system, run into the ground truth of real estate. See the method in full →

01 · Source

Read the authoritative record directly: county assessor valuations, the recorded deeds of trust behind each parcel, the municipal meeting agendas where the next rezone is decided, and the statewide zoning layers. Not scraped listings.

02 · Resolve

Reconcile parcel, ownership as the grantee of the most recent deed, recorded debt with reconveyed and refinanced loans removed, zoning, and site constraints into one clean picture, each item source-cited and any gap left null rather than guessed.

03 · Decide

Score the economics of underutilization, where the dirt is worth more than the building, stack the factual conditions that make a site move now, an opportunity zone, rail, a dated power-grid unlock, a funded interchange, permit momentum, and rank to your buy-box.

04 · Act

Route the decision to the right desk and carry it into outreach, with the owner of record resolved to a dialable contact, and write the result back so the picture sharpens with every deal.

05 · Prove

Every score is reproducible and hash-stamped, every claim traces to the recorded document, and the whole thing was validated out-of-time, with what failed validation removed.

What we tested, and would not ship

Most real-estate AI sells a model that predicts who will sell. We built that model and tested it honestly, on a training cohort and a disjoint later cohort, with every feature built only from records dated before the decision. It did not work. So we do not ship it.

0.55 AUC
the sale-prediction model, validated out-of-time on a disjoint cohort, near-random
0.98×
recorded loan maturity as a sale signal, near-random; owners refinance, they do not sell
Dropped
both removed at zero weight; the score rebuilt on what survived validation
Out-of-time validatedLook-ahead controlledFalsified and removed
What survived is an economic screen and a set of forcing conditions, not a prediction of intent.

The score, not the story

The part a principal cares about most, and the part most real-estate AI cannot answer.

Every score is reproducible and hash-stamped.Deterministic, with no black box. The same parcel and the same records return the same grade, and every run records a hash of its inputs.
Predictions are frozen and tamper-evident.Each score is committed to a hash-chained ledger and checked forward against the record, so the track record is auditable rather than asserted.
Nothing is fabricated.A signal that cannot be verified degrades to null and is labeled as such. And it reads the authoritative record directly, the recorded deeds and civic filings behind each parcel, not a resold aggregator feed that a competitor could license tomorrow.

By the numbers

Built on the public record, at a scale a deal team cannot hold by hand. Every figure reproduces on demand.

40,000+
tenants placed to 18,000+ buildings across 17 counties
4,800+
industrial parcels underwritten across a 20-jurisdiction corridor
22
municipal zoning layers crosswalked to true zoning, not a class proxy
436
predictions frozen and hash-chained, 368 tracked forward to outcome
Deed-level
reads the recorded documents behind each parcel, the depth aggregator feeds never reach
Daily
municipal agendas harvested, so you see the deal before it is listed

The desk that reads the county every morning

Groma values a market and keeps reading. Before the market opens, it has harvested the municipal agendas and minutes across the corridor, extracted every new rezone, annexation, and development agreement with the source excerpt, and written it to an append-only feed. You see the deal at the zoning tell, before it reaches a listing or even the GIS map catches up.

governing bodies readthe corridor · daily
agendas & minutesevery new document, OCR'd if scanned
actions extractedrezone / annexation / dev agreement
every recordsource-cited, append-only
surfacedthe zoning tell, before GIS updates
Illustrative of the daily run. Specific parcels and owners are not shown.

See it run on your market.

ICHOR deploys one industry at a time, forward-deployed and tied to outcomes. Every inquiry is read personally.

Request access