🛠️ From testing to monitoring

We test, correct, and monitor AI systems for fairness, transparency, and robustness—using scientifically validated methods and production-ready tools.

1. Testing (analysis & diagnosis)

We test for …

• Fairness & bias (70+ metrics) in algorithms and training data
• Explainability (why does the system make decisions?)
• Data quality (is training data representative, complete, error-free?)
• Robustness (how does the system react to unexpected inputs?)

How do we test?

• Quantitative metrics (disparate impact, equal opportunity, etc.)
• Qualitative analyses (root cause for bias)
• Subgroup analyses (intersectional fairness)
• Benchmark against best practices

Output:

Detailed test reports with specific findings, severity ratings, and recommendations

2. Mitigation (correction & optimization)

Bias found – what now?

Pre-processing (data):

• Data reweighting (giving greater weight to disadvantaged groups)
• Sampling strategies (balancing)
• Feature engineering (removing/adjusting problematic features)

In-Processing (Training):

• Fairness constraints (e.g., “Demographic Parity ≥ 0.80”)
• Adversarial debiasing (the model learns to be fair)
• Optimization for multiple objectives such as accuracy and fairness

Post-processing (outputs):

• Threshold optimization (different thresholds per group)
• Calibration (output calibration for fairness)

Trade-off analyses:
Every mitigation measure has a cost (usually accuracy). We present all options with clear trade-offs.

3. Monitoring & Reporting (Continuous Monitoring)

AI systems change during use – monitoring is mandatory (EU AI Act Art. 72).

What do we monitor?

• Data drift (are input data changing?)
• Concept Drift (are patterns changing?)
• Performance drift (is accuracy declining?)
• Fairness drift (does fairness deteriorate over time?)

How do we monitor?

• Automated monitoring pipelines (Alibi Detect)
• Real-time alerting (for critical thresholds)
• Dashboard visualization (Evidently AI)
• Quarterly Reports (for compliance verification)

Output:
Live dashboards, alerts, quarterly compliance reports

💡 We cover the entire lifecycle: from initial diagnosis to correction and continuous monitoring.

⚖️ Comprehensive regulatory expertise

We understand not only WHAT the EU AI Act requires, but also WHY, and how to implement it in practice. From fundamental ethical principles to audit-proof documentation.

1. Derivation—Why is regulation necessary?

We understand the philosophical and ethical foundations:

• EU Charter of Fundamental Rights (dignity, non-discrimination, data protection)
• OECD AI Principles (transparency, fairness, accountability)
• High-Level Expert Group Ethics Guidelines (Trustworthy AI)
• Political goals (European Digital Strategy, Green Deal)

Why is this important?
Only those who understand the derivation can anticipate new requirements and interpret them meaningfully.

Example:
“Why does Article 10 require ‘representative’ data?”

→ Derivation: Fundamental right to non-discrimination + OECD Fairness Principle

→ Understanding: It’s about avoiding group bias through skewed training data.

→ Implementation: Distribution matching against target population

2. Objective (What does the legislator want to achieve?)

We know the intentions behind every article:

Key objectives of the EU AI Act:

• Risk-based approach (different requirements depending on risk)
• Harmonization (single EU internal market for AI)
• Protection of fundamental rights (dignity, discrimination, data protection)
• Promotion of innovation (through legal certainty)
• Human Oversight (humans remain in control)

Example articles:
Art. 14 (Human Oversight) → Goal: People can understand and override system
output Art. 72 (Post-Market Monitoring) → Goal: Continuous improvement & risk management

Why is this important?
Understanding the goal enables pragmatic solutions instead of ineffective box-ticking compliance.

3. Implementation (How do you meet requirements technically?)

We translate legal texts into specific technical requirements:

Translation methodology:

Legal text → Interpretation → Technical requirements → Measurable metrics → Tools/processes

Example Art. 15 (Accuracy, Robustness, Cybersecurity):

“Accuracy”:
→ Performance metrics across all subgroups
→ Tool: Fairness Indicators (Subgroup Performance Matrix)
→ Threshold: Max 5% performance gap between groups

Robustness:
→ Adversarial testing (how does the system react to manipulated inputs?)
→ Tool: IBM ART (Adversarial Robustness Toolbox)
→ Threshold value: <10% performance drop for standard attacks

Cybersecurity:
→ Security testing (can the system be hacked?)
→ Tool: Penetration testing + security audit
→ Standard: OWASP Top 10 for ML

4. Use (How can this be implemented in the company?)

From compliance requirements to operational practice:

Governance structures:

• AI Governance Board (who makes the decisions?)
• Roles & Responsibilities (who is responsible for what?)
• Escalation paths (for incidents)
• Review cycles (how often to review?)

Processes:

• AI System Lifecycle Management
• Risk assessment workflows
• Incident Response Procedures
• Documentation Standards

Integration:

• Integration into existing processes (GDPR, ISO 27001, etc.)
• Toolchain integration (CI/CD pipelines)
• Training & Awareness (Empowering teams)

Example:
Art. 9 (Risk Management System):
→ Governance: Risk Committee (monthly meetings)
→ Process: Risk assessment for each release
→ Integration: Risk scores in JIRA tickets
→ Training: Quarterly risk assessment workshops

💡 We think from principle to practice: Why → What → How → Who/When/Where