Five situations, one movement

IThe interview corpus

Forty transcribed interviews, months of fieldwork, and one conviction: a public AI would make a formidable reader — coding themes, spotting tensions, suggesting leads. But every page carries names, employers, towns. Sending them out as they are would betray the promise made to the people interviewed.

The movement: the whole folder goes through the airlock. A single table: each interviewee keeps the same false name from interview 1 to interview 40 — cross-cutting analysis remains possible. The corpus goes out, the AI works, the syntheses come back, and restoration puts the real names back — including in the texts the AI has rewritten. The originals, for their part, never moved.

IIThe DPO who has to be able to explain

A department wants to use a public AI on internal documents. The data protection officer's question is not “is it magic?” but “can I explain and defend this processing?”.

The movement: Pseudo is explainable end to end. The list of names is validated by a human, one by one; a real name can only stay in clear text on a reasoned decision; the final scan refuses to write a copy as long as a name from the list survives in it; the table is encrypted, and the key stays in-house. And the vocabulary is honest: pseudonymisation, in the GDPR sense, not a bolder word — the limits (cross-referencing, human review) are stated in the application itself.

IIIThe client file that needs a second reading

A consultant, a lawyer, a chartered accountant: the deliverable would benefit from an AI review, but the file carries SIRET numbers, IBANs, phone numbers, and directors' names.

The movement: French patterns are detected and replaced with valid substitutes — an IBAN whose check digits add up, a number within the official fictitious ranges — so the document stays plausible and workable. On the way back, each pattern returns exactly as written. The client was never exposed.

IVThe full round trip, step by step

The classic doubt: “what if the AI rewrites everything — how do I get my names back?”. This is the heart of the product: restoration works on any text that carries the pseudonyms — an exact copy or a document reworked from top to bottom. Send the file out, have it rewritten, restore: the real names come back into the rewritten text. When a passage is ambiguous, the app would rather show it to you than guess.

VThe list review as the central scene

A surname that is also a common word; two namesakes; an organisation cited under three spellings. No automaton settles that cleanly — and Pseudo does not claim to.

The movement: the app proposes — a list, doubts set apart, excerpts in support — and you decide: merge, remove, add a spelling, rule “common word here, identity there”. It is slower than a magic button, and that is exactly why you can put your name to the result.