From dense documents to data you can audit.

Source material is messy by nature: installation manuals, spec sheets, nomenclature tables, and product APIs that were never designed to be read by machines.

Documents move through a change-tracked lifecycle (cataloged, downloaded, promoted, extracted) with dated snapshots and diff reports, so the platform knows when a manufacturer quietly revises a publication. Product APIs are mapped exhaustively: parameter spaces enumerated configuration by configuration, with convergence probes that prove the enumeration is complete rather than assuming it.

A multi-stage Python pipeline turns PDFs into structured claims: table geometry, spec values, units, conditions, and the nomenclature systems that encode a model family into its model numbers.

Every stage emits observable intermediate outputs and findings reports. Nothing is a black box: when a value looks wrong downstream, the pipeline can show exactly which page, table, and cell it came from.

Extraction without verification is guessing. Pipeline outputs are checked against independent authoritative product indexes: ground truth the extraction cannot see during processing.

Discrepancies get a three-verdict ruling: match, variant, or drift. Human adjudications persist in a ledger, so a decision is made once and reapplied automatically. A promotion gate enforces the contract: unverified data does not ship.

Verified claims land in an OLAP fact-table schema of equipment identities, spec facts, adjustment rules, and provenance: normalized, dimensional, and built for querying rather than display.

Every fact row has a provenance row: source document, page, table, extraction method, confidence. Provenance is not metadata bolted on afterward; it is carried through from the first parse.

The same fact tables serve three surfaces: a REST API with per-field provenance, an MCP server that makes the dataset directly queryable by AI agents and chat tools, and versioned columnar Parquet snapshots for teams that want the data itself.

Human review tooling closes the loop: a PDF-overlay interface where extraction results are inspected against the source document, and every ruling feeds back into the ledger.