National Technical University of Athens (NTUA)
Trustworthy AI operations with conversational assistance and agentic workflows
Abstract
dc:descriptionThe rapid advancement of Artificial Intelligence (AI) capabilities—from static models to autonomous agents—has created a critical "Operations Gap," where the ability to develop powerful models has outpaced the capacity to integrate, monitor, and control them reliably in real-world environments. Traditional operational paradigms like Machine Learning Operations (MLOps) are often insufficient for modern agents that exhibit autonomy, non-determinism, and direct interaction with humans. This thesis addresses this gap by establishing the discipline of Trustworthy AI Operations, providing a roadmap for deploying autonomous AI that is not only powerful but also reliable, transparent, and aligned with human values.A central contribution of this work is a novel taxonomy that characterises the operational maturity of AI systems through five evolving levels: Handcrafted, Integrated, Augmented, Trusted, and Autonomous. Handcrafted systems rely on manual, disjoint implementations with minimal automation. The Integrated level unifies lifecycle management, where data, model training, and deployment are orchestrated within a cohesive architecture to eliminate silos. The Augmented level enhances human-AI interaction using Large Language Models (LLMs) and intelligent agents to lower the cognitive barrier of managing complex systems. The Trusted level operationalises trustworthiness, ensuring requirements such as fairness and explainability are continuously verified. Finally, the Autonomous level represents the future state of self-governing multi-agent systems operating within established trust boundaries.To support this evolution, the thesis introduces a reference architecture that unifies disparate AI development stages, enabling automated workflows and reducing technical debt. It explores agentic augmentation, where conversational AI acts as an intermediary between human intent and system operations, democratizing access to complex analytics. Furthermore, it defines the Trustworthiness Optimisation Process (TOP), a continuous loop of identification, assessment, exploration, and enhancement that treats sociotechnical constraints as real-time operational objectives rather than one-off compliance checks.As a technical validation of these concepts, the thesis presents Trustproofer, a neuro-symbolic agentic framework. Trustproofer combines the reasoning capabilities of symbolic AI with the adaptability of neural networks to orchestrate trust-aware operations. It serves as a practical testbed, demonstrating that high-level autonomy can coexist with rigorous human oversight and sociotechnical safeguards, thus bridging the gap between theoretical AI capabilities and their safe, effective deployment in production.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Fikardos, Mathaios
- Φικάρδος, Ματθαίος
Subjects
dc:subject × 13Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61226