TRARDI Group: remit, organisation and operating principles

TRARDI Group works on the examination, design and control of artificial-intelligence systems. Each engagement specifies the scope of work, the responsibilities of those involved and the evidence required. Conclusions set out the findings, their limitations and the measures to be considered. The transfer of the knowledge and documentation required for operation forms part of the engagement scope.

In practice this means four bounded engagements: arbitrate a portfolio of use cases before funding, audit a system before committing, measure a dependency before it becomes irreversible, and verify that a system in production still delivers the approved service. Scope and timetable are agreed during scoping; each engagement ends with a documented decision.

Youssef Trardi completed a doctorate at Aix-Marseille University in 2020, specialising in machine learning, analysis and diagnosis. He later led CEDRE, its programme for research data and artificial intelligence. He is the founder and managing director of TRARDI Group.

Depending on the mandate, he brings in named practitioners in audit, architecture, cybersecurity and data. Their names, roles and availability appear in the engagement letter; no subcontractor works on an engagement without being identified to the client.

If a design cannot be explained to the people who will depend on it, we do not consider it deliverable.

Dr Youssef Trardi

It requires neither building everything yourself nor excluding suppliers. It requires understanding what decides inside the system, choosing its dependencies, controlling its data and access, being able to operate in degraded mode, and being able to leave a provider without losing service continuity. Everything we deliver serves that capability: it is observed in practice, never on a certificate.

These commitments decide a scope, an architecture, a control or the closure of an engagement. If they changed no decision, they would be statements of intent.

• Purpose is written before the solution: you do not fund a tool, you inform a decision, and the no-AI scenario stays in the comparison.
• A claim is worth the evidence behind it: every conclusion states the criterion, the version examined, the procedure and what could not be verified.
• The owner must be able to decide: naming someone without access, budget or the power to stop the system creates no accountability.
• Operations are designed before release: monitoring, incidents, degraded mode and rollback are part of acceptance, not of maintenance.
• Reversibility is an exercise: an exit clause is not enough; we export, restore and time it.
• Transfer is observed in practice — the engagement closes when your team handles a real deviation without us, not when the training ends.

We are neither a certification body, nor a law firm, nor a systems integrator. Our work produces methodologically defensible conclusions, with written reservations; it does not replace an official conformity assessment, legal advice or a competent authority. The mandate declares commercial relationships and conflicts of interest that could affect the examination.