How is TRARDI different from a traditional consultancy?

Traditional firms stop at the strategy deck. TRARDI designs, builds, deploys, and measures adoption — same team, end to end. We also run 3 proprietary multi-agent AI platforms (Talent360, SyncAgent Pro, Praxis), so the discipline we teach is the one we practice daily.

Can we start small before committing?

Yes. Most engagements start with a 1–2 week fixed-price AI Readiness Audit. You get a scored roadmap with go/no-go per use case and the TRARDI Impact Score™. No follow-on commitment required.

Do you handle on-premise, sensitive data, and GDPR?

Yes. We design on-premise, hybrid, or private-cloud AI environments for regulated industries (finance, health, defense, public sector). Every engagement includes a signed DPA. TRARDI GROUP is a French SARL based in Paris (RCS 102 266 814).

What happens after deployment?

A mandatory adoption program — we don't leave systems unused. We measure real adoption with the Adoption Velocity Index™ (active users, task completion rate, time-to-autonomy). Then an optional governance retainer transfers ownership to your team.

What AI models and infrastructure do you use?

Frontier models only — OpenAI, Anthropic, and custom fine-tuned stacks. We never compromise on model quality. Infrastructure runs on your stack: Kubernetes, vector stores, and a fully auditable MLOps pipeline.

Who owns the code and models?

You do. Every engagement ships with full IP transfer: source code, model weights, configuration, documentation, and operational playbook. TRARDI is not a SaaS vendor — we build systems that run without us.

TRARDI Framework · Part 02

AI Systems Audit

Is this system built correctly?

AI Systems Audit: six-dimension review of a live AI system covering context, architecture, model, controls, supervision, and governance. Severity-ranked findings with named owners. Version 0, 2026.

What an AI Systems Audit covers

An AI system is more than a model. An AI Systems Audit assesses the system as a system: the architecture around the model, the data it depends on, the controls that constrain it, the supervision that catches its errors, the journalization that makes it reviewable, and the governance that owns it. The output is a findings report with severity-ranked gaps and named risk owners for every high or critical finding.

The six audit dimensions

Each dimension is rated 1 to 5 on the framework scoring scale. Composite scores roll up to 0 to 5 for the system.

Context & classification

What the system is for, which risk class applies under the AI Act, which regulatory obligations are triggered.

Key checks

Intended purpose documented and signed
AI Act risk class determined (minimal, limited, high-risk, or prohibited)
Regulatory obligations mapped to operational controls
Affected parties identified (users, subjects, third parties)

Architecture & data

System design, deployment topology, data flows, dependency graph, integration boundaries.

Key checks

Architecture diagram current and matches production
Data lineage traceable from source to output
Third-party dependencies listed with update cadence
Integration points documented with failure modes

Model & pipeline

Training data, validation regime, deployment pipeline, versioning, rollback procedure.

Key checks

Training data sources identified and license-cleared
Validation protocol reproducible
Model versions tagged with deployment dates
Rollback path tested within the last quarter

Controls

Access control, input validation, output filtering, rate limiting, safety mechanisms, policy enforcement.

Key checks

Access control enforced at API and infrastructure level
Input validation rules documented and tested
Output filtering prevents policy violations
Rate limiting configured against denial-of-service

Supervision & human oversight

Human-in-the-loop checkpoints, escalation procedures, override paths, oversight cadence.

Key checks

Human checkpoints defined for high-risk decisions
Escalation rules trigger automatically on defined conditions
Override actions logged with operator identity and reason
Supervision team staffing sufficient for decision volume

Evidence & governance

Audit trail, documentation maturity, incident response, named ownership, review cadence.

Key checks

Audit trail complete, immutable, retained per policy
Operating documentation current within the last quarter
Incident response procedure defined, tested, and named
System owner named with decision authority documented

Auto-block rules

Four conditions trigger automatic production sign-off block, regardless of composite score.

No named system owner

Block

An unowned system cannot be governed.

No audit trail

Block

A system without logs cannot be audited after the fact.

No incident response procedure

Block

An incident without a procedure becomes an improvisation.

Output policy filters untested

Block

A system that can violate policy in production cannot ship.

What you receive

An AI Systems Audit produces six concrete deliverables, handed over at the end of the engagement.

01
Audit scope memo with signed boundaries
02
Evidence pack organized per dimension
03
Findings report with severity-ranked gaps
04
Remediation roadmap with named owners and deadlines
05
Executive debrief (60 minutes, recorded)
06
Re-test plan for high and critical findings

Typical duration

Two to four weeks depending on system scope. Week 1: scope and evidence collection. Week 2: dimension review and scoring. Week 3 to 4: findings report, debrief, remediation roadmap.

Scope of this method

This method produces a private audit report for the client organization. It is not a certification under ISO/IEC 17021, not an AI Act conformity assessment, and not a substitute for official regulatory audit by accredited bodies. It is the discipline TRARDI applies to structure its assurance opinions.

Want an audit of one of your AI systems?

Book a 30-minute diagnostic

We will scope the audit to a specific system in your portfolio and walk you through which dimensions would apply first. No pitch.

Book a diagnostic