How it works

Keltes is being built to read your documents first and to ask only what they cannot answer. Today the layer that reads documents has run only on fictional documents and its quality is not yet measured.

AI reads your documents, links passages to obligations and drafts changes and reports. It never decides what applies, never produces a score and never marks anything as final: fixed rules and people do. Every proposal carries its article, the passages it relies on, the model, the prompt version and a hash of the input. When an assessment is reproduced, the recorded AI outputs are replayed, never generated again.

Article 18(3) of the AMLR lists six tasks that shall not be outsourced, among them the approval of the business-wide risk assessment and of internal policies, procedures and controls, and reporting to the FIU, with narrow exceptions within a group and for some funds (Article 18(3), point (e), and Article 18(7)). Keltes holds the six as a register. Every recommendation names the role in your firm that decides, and a test fails if any part of the engine defines a decision type.

The rules are written down once, as data, in a dated release, and every firm is assessed against the same release. A release can be used for a firm only once the founder has approved it and signed it with his key, and only while it passes every automatic check. A live release is never edited in place: a correction goes into a new version, with a record of what changed.

Every output says how its rules were checked, for example: "Rules checked by AI and approved by the founder; not reviewed by an independent practitioner." A practitioner's review is an optional record; the output mentions one only where it was bound to the release, and never claims more. A known gap in what the release covers is named on the output as not covered.

No release has been approved or signed yet.