As Artificial Intelligence (AI) systems move from prototypes to deployment, reliability and trustworthiness are increasingly limited by the data and AI pipeline lifecycle rather than by model training alone. Existing platforms offer strong support for individual lifecycle functions such as orchestration, experiment tracking, validation, or documentation, yet these capabilities are often adopted as separate tools, making provenance, observability, governance, and safe adaptation difficult to manage end-to-end. This paper presents the microservice-based reference architecture of the AI-DAPT project for reproducible AI pipelines whose core contribution is the unification of data- and model-centric operations within a single lifecycle-oriented platform. The architecture is organized around five iterative phases, namely data design, data sculpting, data generation, model delivery, and data/model optimization, while treating metadata, lineage, explainability, security, and human-oversight as native platform services rather than peripheral add-ons. The paper further contributes a trustworthiness-by-design architectural view that connects data preparation, valuation, synthetic data generation, orchestration, monitoring, adaptive retraining, and AI security through interoperable services. Finally, it presents the current realization of this architecture using widely adopted orchestration, storage, monitoring, and experimentation technologies, thereby offering a practical blueprint for building adaptable, observable, and governable AI pipelines in real-world settings.

Agentic AI systems are increasingly deployed within complex pipelines, where failures often arise from interactions across reasoning steps, tool calls, and external environments. In such settings, understanding system behavior requires diagnostics grounded in execution traces rather than post-hoc, high-level explanations. To assess this requirement, we study trace-aligned diagnostics for monitoring and debugging agentic AI pipelines. Using the TRAIL benchmark, we implement a lightweight span-level model that identifies and localizes failures, and compare its performance to LLM-based baselines. Our results show that trace-aligned methods outperform LLM-based approaches (0.48 vs. 0.39 category F1 and 0.61 vs. 0.55 localization accuracy on GAIA) without suffering from context-length issues, highlighting the importance of execution-grounded analysis for trustworthy AI systems.

The growing adoption of AI/ML and data analytics pipelines in industrial and research settings has increased the demand for accessible, user-centered tools for pipeline management. While existing orchestration platforms provide mechanisms for execution tracking, they often expose complex execution information in ways that are difficult to interpret, limiting their usability. This paper presents the Data Analytics and Visualization Environment (DAVE) Execution Environment, a user-centered interface designed to support the execution, monitoring, and debugging AI/ML and ETL pipelines within the DAVE framework. The proposed approach is grounded in principles of visual clarity, task-level observability, and progressive information disclosure, enabling users to interpret pipeline behaviour and diagnose issues more effectively. The solution has been validated across multiple pilots in different sectors, demonstrating its effectiveness in improving the accessibility and interpretability of pipeline execution for non-expert users. These results highlight the potential of combining visual representations with task-level inspection to bridge the gap between pipeline complexity and usability in no-code/low-code MLOps environments.