Open-source intelligence has a reputation for being slow work done by hand. An analyst opens twenty tabs, reads for an afternoon, and writes a memo that the next analyst cannot reproduce. The findings may be right. The process is not something a bank can put inside an onboarding flow.
The pipeline exists to make that same work repeatable, fast and reviewable. Four stages, always in the same order.
Collection
The query fans out to more than five hundred public sources at once: company and public registries, published press, open web data and public profiles. Every source is public by law. Nothing in the pipeline asks for access it was not given.
Breadth is the point. A single match is a coincidence. The same person appearing in a registry, a court filing and a news item is a pattern, and patterns are what the next stage reads.
Correlation
The engine links what the sources say about each other, not just what each one says alone. Two records with the same name and different dates of birth are two people. Two records with different transliterations of a name, the same date of birth and the same city are probably one.
This is the part that used to take the afternoon, and it is the part a machine is genuinely better at. It never gets bored on the four hundredth record.
Scoring
Every query ends in a risk score from 0 to 1. The score is not a verdict. It is a reading instruction: the higher it climbs, the more the report asks a person to read before anyone decides.
A score you cannot open is a rumour. Every number in the report traces back to the source that produced it.
Delivery
The report comes back as JSON on the same API as document and biometric verification, and as a readable document for whoever has to sign off. Both carry the same sources, in the same order, with the same timestamps.
That is what makes the result defensible six months later, in front of a regulator who was not in the room when the decision was made.