Answers
Questions we are actually asked
One question per page, answered in the first block, with the reasoning and the primary sources underneath. They are written from what this practice has built and had to defend — where an answer would require inventing evidence, the page does not exist.
What is the difference between private AI and sovereign AI?
Private AI is about who can see the data. Sovereign AI is about who can decide, and under which jurisdiction. The three deployment models — fully local, sovereign hybrid, EU private inference — and what each one actually guarantees.
Is a RAG system subject to the EU AI Act?
The AI Act classifies by intended purpose and by role, not by technique. What that means for a retrieval-augmented generation system, which obligations attach to the deployer rather than the model provider, and what the architecture has to be able to show.
How do you build a RAG system with an audit trail?
What has to be recorded for an answer to be reconstructable months later: retrieval set, document versions, prompt, model identity, and the approval. The design decisions that make it possible, and the ones that make it impossible.
How do you handle superseded regulatory documents in a RAG system?
A retrieval system that ranks by similarity will happily return a circular that was revoked two years ago. Version chains, validity intervals and the difference between a document being relevant and being in force.
How do you evaluate an AI vendor in a regulated sector?
The questions that separate a supplier who has thought about accountability from one who has not: data flows, subprocessors, model substitution, audit surface, incident duties and where liability actually stops.
How do you implement an LLM on-premise in a hospital?
Sequence, hardware, data boundary and governance for a hospital deployment: what to start with, what personal data should never reach the index, and which parts of the stack decide whether the project survives its first audit.
Longer arguments live in Research; what a deployment looks like in production is in the case studies. If your question is not here, it is worth asking directly.