AI systems your board can defend, and your teams can actually run.

Tell us about your situation See our engagements›

Three moments when an independent view changes the decision

Before you sign, before you go live, and while the system is already running.

A supplier promises performance, sovereignty and an easy exit. We test all three before you commit.

  • Claimed performance
  • Real dependencies
  • Exit clauses
  • Qualify
  • Arbitrate
  • Architect
  • Control
  • Document
  • Transfer

Our engagements

Four engagements, one for each stage of the system.

Each answers one precise question and ends with a file your leadership can read and your teams can maintain.

Before funding

AI portfolio arbitration

Compare use cases across five documented axes and address the conditions that require a stop or deferral.

Scored portfolio · documented decisions

Before committing

AI systems audit

Verify a system as it actually operates: architecture, data, models, controls, usage and dependencies.

Findings · severity · treatment plan

Before depending

Reversibility audit

Measure what you recover if a supplier stops, and run a real, timed exit test.

Dependencies · timed exit test

In production

AI production assurance

Verify that the system still delivers the approved service, and that every drift leads to a decision.

Baseline · drift · operating decision

Compare the four engagements

1 / 4 — AI portfolio arbitration

EU AI Act · Regulation (EU) 2024/1689

From obligations to evidence.

We qualify the system and your organisation’s role, then connect each requirement to an owner, a control and evidence you can actually present.

  1. Actual system use Purpose and context
  2. Organisation’s role Provider or deployer
  3. Risk level Reasoned classification
  4. Applicable requirements Duties and timeline
  5. Controls and evidence Owners and traceability
  6. Audit-ready file Readable by leadership, maintainable by teams

Generally applicable since . High risk: Annex III on ; systems embedded in regulated products on . Official source

Read our AI Act page›

How an engagement runs

Five steps, a named owner at each.

  1. Mandate

    You give us a decision to inform, a deadline and the authority who will decide.

  2. Examination

    We examine the system in its real version and test statements against traces.

  3. Decision

    Each finding gets a severity rating, an owner and an expected decision.

  4. Release

    Your leadership authorises, restricts, defers or refuses on an identifiable version.

  5. Transfer

    The engagement closes when your team can qualify an alert and decide a fallback without us.

How we work›

The group

Scientific direction and responsibility for engagements

TRARDI Group is led by Dr Youssef Trardi, who earned a doctorate from Aix-Marseille University in machine learning, analysis and diagnosis, and later led the CEDRE programme for research data and AI.

Discover the group›

First conversation

Which system or decision do you need to secure?

Describe the system, the deadline and the main risk, without sensitive data. We reformulate the question and, where an engagement is relevant, propose a written scope.

Tell us about your situation