DAYA: Data-Aware Agentic AI Architecture for Decision-Oriented Analytics on HPC Systems [09:55 - 10:20]

Authors: Michael Bidollahkhani, Orr Shomroni, Hassan Dihazi, Julian Kunkel

Scientific and medical data analytics increasingly depend on complex computational workflows that combine heterogeneous datasets, domain knowledge, machine learning models, and interpretation-oriented decision processes. In high-performance computing (HPC) environments, these workflows often involve large-scale data processing, repeated model execution, simulation outputs, and computationally demanding analysis pipelines. As a result, analysts face challenges related to data understanding, method selection, computational cost, result interpretation, and the translation of analytical outputs into reliable decisions.
This talk presents DAYA, a Data-Aware Agentic AI Architecture for Decision-Oriented Analytics on HPC Systems. DAYA is motivated by practical challenges in real-world data analysis, especially in medical and scientific domains, where users need to understand the structure, quality, and meaning of data before selecting suitable analytical methods. The central idea of DAYA is to combine the roles of a data specialist and a data analyzer within one agentic AI architecture. Instead of applying a fixed analysis pipeline to every dataset, DAYA reasons about the available data, analytical goal, computational context, and expected decision outcome before selecting an appropriate analysis path.
Within the context of HPC-driven decision science, DAYA is positioned as an adaptive agentic AI layer for data-aware decision support. The architecture is designed to coordinate multiple stages of the analytics process, including data inspection, quality assessment, method recommendation, model selection, analysis execution, and result interpretation. This allows DAYA to support analytical workflows where the final decision depends on both computational results and contextual understanding of the data. Such a design is particularly relevant for HPC-based analytics, where large datasets, simulation-related outputs, and repeated analytical tasks require scalable and systematic AI-supported methods.
The proposed talk is structured in two parts. First, we discuss the problem statement from the perspective of real-world data analysis, with emphasis on medical data, data heterogeneity, analytical uncertainty, and the difficulty of transforming raw data into decision-ready knowledge. Second, we present the DAYA architecture as a solution that combines data-aware reasoning, agentic workflow control, and decision-oriented analytics. We also discuss OUTCAST AI as a planned use case for demonstrating how DAYA can support adaptive analytics and AI-assisted decision-making in practice.

Session Chair: Andreas Fink

15.09.2026 09:30 - 10:50 Lecture Hall 3
Contributed ‧ HPC-Driven Decision Science