How it works
A title search package in. A structured exam file out.
Here is what happens in between.
Upload the search package.
One PDF with the recorded documents you pulled from the county. Deeds, mortgages, releases, assignments, liens, lis pendens, tax bills, judgments, UCCs, court orders. Whatever your search returned. Could be five pages, could be two hundred.
We accept the package as-is. No pre-segmenting required. No filename conventions required.
Each recorded instrument gets pulled out.
The engine reads the OCR text of the whole bundle and identifies where each instrument starts and ends. A 90-page bundle might decompose into 18 separate instruments. Each one gets its own card in the final report.
The segmentation knows the patterns: page-of-N markers, "END OF DOCUMENT" footers, recorder stamps, header changes. When it can't tell where a boundary lives, it flags the page for examiner review rather than guessing.
Each instrument is extracted against a typed schema.
Per-document-type extraction prompts. Deeds go through the deed prompt. Mortgages go through the mortgage prompt. Lis pendens, NFTLs, hospital liens, mechanics liens, UCC filings. Sixteen different prompts in total, each tuned to that instrument type.
Every field the engine emits is type-checked against a Pydantic schema. If the model hallucinates a field, the schema rejects it. If the model omits something required, the document goes to needs-review. No silent drift between what the prompt asks for and what the consumer expects.
The chain gets read like an examiner would read it.
Once the documents are extracted, the engine walks the chain the same way you would. It checks the things you check, and surfaces what it finds for you to confirm.
Foreclosure that went all the way to sale
When a lis pendens and a certificate of title share a case number, that's a completed foreclosure. The engine pairs them up, points at the mortgage that got foreclosed, and asks you to confirm which junior liens went with it.
IRS lien still in its redemption window
If a federal tax lien was on title before a foreclosure sale, the IRS has 120 days to redeem. The engine does the date math and tells you whether the window is still open. It does not tell you whether the IRS was actually notified. That's still your read.
Assignment chains, including the MERS dance
The engine follows the assignment chain from origination through MERS to whichever servicer or trustee actually holds the note today. It tells you who to ask for the payoff letter.
Dissolved-entity grantors
If a deed comes from an LLC that's already been dissolved by the secretary of state on the recording date, that's a flag, not a fact-pattern you usually catch by eye. The engine surfaces the dissolution date so you can decide whether it matters here.
And the rest of the catalogue
Tenancy-by-the-entirety in the 21 states that recognize it. F/k/a, a/k/a, n/k/a name-change bridges so the chain doesn't break on a marriage. Re-record pairing on attribute match. Open mortgages without a corresponding release. Forty-plus checks like this, each one written from a real package we got wrong the first time.
The rule for every check: the engine surfaces the fact, the examiner makes the call.
The rendered exam file.
Four sections in the order examiners expect: Deeds and Conveyances, Mortgages, Judgments and Liens, Taxes and Assessments. Each parent instrument shows its dependent trailings nested below: a mortgage with its release, its assignment, its modification, its ALR.
Every callout in the report cites the document and the page it came from. The examiner opens the PDF, reads the engine's observations, verifies against the source pages, signs off.
Per-order transparency, your pricing.
Each order returns a per-order cost breakdown: OCR pages, model tokens, render time, so you can see what drove the cost.
Pricing to you adapts to volume, product type, and the work you're sending. We'll talk through it directly rather than pretend one rate fits every shop.