The Pipeline
Seven stages, fully autonomous.
Each stage is independent, fault-tolerant, and resumable. Notices are re-read and re-arbitrated several times over the cycle — the system runs entirely without human intervention.
01
Continuous monitoring across all 254 counties
Custom automation agents - one per county clerk platform - run on a scheduled cycle, navigate each site, and download every newly-filed Notice of Trustee's Sale. The dispatch system selects the right automation agent for each county based on the platform it runs.
Custom Automation Agents15+ Platforms254 CountiesMonthly Cycle
02
Two independent LLMs read every document
Each notice is processed by two separate language models with independent prompts and independent output schemas. They don't see each other's output. This deliberate redundancy means a hallucination or transcription error in one model can be caught by the other - no single point of failure in the most error-prone part of the pipeline.
Analytical LLMHigh Reasoning LLMStructured JSONIndependent Schemas
03
Property geometry, computed from the legal description
Texas legal descriptions carry raw metes-and-bounds courses - bearings, distances, and turn calls that trace the property's perimeter. Texas NTS extracts every course straight from the document, algorithmically normalizes them into a standard format, then runs a real-time area and closure calculation on the parsed geometry. The computed acreage and closure error get packaged as computational evidence and passed to the arbitrator - so any acreage figure the LLMs read from the text is cross-checked against the geometry the property itself implies.
Course ExtractionFormat NormalizationReal-Time ClosureComputational Evidence
04
A third model resolves disagreements field-by-field
When the two extraction models disagree on any field - borrower name, property address, sale date, legal description, acreage - a third arbitration model reviews both candidates against the original OCR'd document plus the computational evidence from the geometry pipeline, and picks the correct one. Every arbitration decision is logged with reasoning. The result: a structured audit trail attached to every record, with text-extracted figures cross-checked against deterministic computation.
Arbitration LLMOCR + Geometry TiebreakerField-Level Audit Trail~3.7% Trigger Rate
05
Re-read across the cycle, refine on every pass
Each notice doesn't get read just once. As the cycle progresses, the pipeline circles back and re-reads every active notice multiple times - each pass running its own pair of independent extractions and its own arbitration. The canonical record updates on every pass, every disagreement is logged, and confidence climbs as evidence accumulates. By sale day, a typical notice has been read and re-arbitrated many times over.
Multi-Pass ReadsPer-Read ArbitrationCompounding AccuracyAudit Trail Per Pass
06
Same property, different listings - collapsed to one record
Foreclosure notices get re-filed, postponed, and re-listed with subtle variations. Texas NTS uses complex proprietary algorithms across borrower names and legal descriptions, with vesting language stripped, to identify when two notices describe the same property. A challenger rule weighted by read count resolves ambiguity. The result: one canonical record per property, with the full filing history attached.
Borrower Name MatchingLegal Description MatchingVesting NormalizationChallenger Rule
07
Structured data, in your dashboard or your stack
Every notice is normalized into a queryable schema - borrower, property address, parcel ID, legal description, computed polygon area, sale date, trustee, confidence score, audit trail. Self-serve customers work in the Notice Explorer with saved searches and daily email alerts on new matches, plus bulk CSV and structured JSON export. Enterprise customers add direct REST API access and custom webhooks for real-time delivery.
Notice ExplorerSaved Searches & Email AlertsBulk CSV / JSON ExportREST API & Webhooks (Enterprise)