AI governance
I design AI governance models that keep advanced analytics and AI-enabled services accountable, reviewable, and safe to operate in public sector and regulated environments.
What you get
- AI use-case inventory and EU AI Act risk classification
- Accountability model and human-in-the-loop design
- Data provenance and quality controls
- Security testing — prompt injection, access, leakage
- Evaluation baselines and drift monitoring
- Audit trail and ISO 42001-style management system
Applied at: PIR2-IT · 35 states · banking and defence clients · Romanian Government & EU programmes
Institutional foundation
AI without governance creates legal, operational, and reputational risk. The objective is not experimentation alone—it is controlled adoption with human oversight, data integrity, evidence capture, and defensible decision pathways.
AI governance experience across institutions and industry
Representative patterns include accountable AI operating models, human-in-the-loop controls, model-risk governance, traceable decision support, and integration of AI into architectures that must remain secure, auditable, and aligned to institutional mandate.
What I Do
Core capabilities: AI governance frameworks, human oversight models, model-risk controls, data integrity rules, auditability requirements, and architecture patterns for safe AI adoption. Mission context: public sector, regulated industries, high-trust institutions, and transformation programs where accountability cannot be outsourced to algorithms. Outcome: AI becomes governable, explainable, and operationally defensible.
Core principles
- Human oversight over consequential decisions
- Traceability of inputs, outputs, and decision logic
- Data integrity and controlled use of sensitive information
- Auditability and evidence retention
- Security, access control, and operational guardrails
Why it matters
AI is only credible when it remains aligned with law, policy, institutional mandate, and executive accountability. In high-trust environments, control design matters as much as model performance.
Impact in numbers
Figures as recorded on named engagements — each one is listed on the Projects page.
How I approach it
- Inventory and classify. Every AI use case listed and risk-classified against the EU AI Act.
- Name an owner. One accountable owner per system, and a human in the loop for consequential decisions.
- Govern the data. Provenance, quality and lawful basis for training and retrieval data.
- Secure the model. Authentication, least-privilege tool access and testing against prompt injection.
- Evaluate continuously. Baselines for accuracy, bias and drift, with thresholds that trigger review.
- Keep the audit trail. Inputs, outputs and decisions logged; an ISO 42001-style management system around it.
Where it was applied
- PIR2-IT · 2024 – PresentAI governance and agentic eJustice across 35 states — NLP case routing, AI-assisted document processing, digital-evidence chain of custody.
- PIR2-IT · 2025 – 2026Phase 2 agentic AI upgrades for eJustice in Armenia and the 112 emergency service in Georgia.
- Banking and defence · 2024 – PresentAgentic AI for IT service operations in ServiceNow — multi-agent orchestration with RAG, human-in-the-loop and audit guardrails.
- Romania / EU · 2020 – 2023AI policy and digital-strategy advisory for the Romanian Government and EU programmes.
- Innovation Center · 2019 – 2022CyberPro — AI-driven cybersecurity platform integrating SIEM, SOAR, EDR and DLP.
Where programmes usually go wrong
- Pilots without an owner, a budget line or an exit criterion.
- Agents given write access to production systems without guardrails.
- No evaluation baseline — nobody can say whether the model got better or worse.
- Prompt injection and data leakage never tested before go-live.
- Documentation written for the regulator, not used by the operators.
Questions worth asking before you fund it
- Which decisions can the AI take without a human — and who signed that off?
- Can you replay why the system gave a specific answer last month?
- Who tested it against prompt injection and data exfiltration?
- Where is the off switch, and who is allowed to use it?