Document extraction and reconciliation for Capital Markets

This agent reads a full institutional onboarding pack, where unlabelled documents are supposed to describe the same people and company but often disagree. Users supply a JSON Schema of the fields they need, and the agent extracts each one with the verbatim quote and page that support it.

It then works out which claims refer to the same person or company, matching name orderings, initials, and legal-entity suffixes, and compares every field across every document that mentions it. Each field is marked agreed, single-source, conflicting, or missing. Conflicting readings are returned side by side with none chosen, single-source values are labeled as such, and a missing field is proven absent by an exhaustive pass rather than inferred from an empty search. Matching is deterministic, so addresses that differ by one digit are caught and one name written three ways is recognized as one person, while the language model only reads documents and writes the summary. The agent never marks a fact as verified; it gives a reviewer the evidence, disagreements, and gaps.