This workshop explores these challenges through real-world applications examples.

Abstract

How can agentic AI be brought to the extreme edge, where IoT nodes are not connected to the cloud for either practical or cybersecurityreasons and do not have the computational power to run local models?

How can a chatbot be deployed in the medical domain and guarantee that, when it makes mistakes or hallucinates, the user is made aware of that uncertainty?

How canhighly specialized skills be taught to the agentic system? How canagentic AI be used not only to support experts but also to learn from and with them?

How can an agentic pipeline be evaluated and validated in the context of highly specific data, such as time series, where academic datasets and existing benchmarksare unsuitable? How can performance be reliably benchmarked across different models?

This workshop explores how CSEM brings agentic AI from research to real-world use, addressing edge deployment, medical trustworthiness, industrial time-series reasoning and expert-driven microscopy workflows.

Workshop description

Agentic AI is moving beyond general-purpose chatbots toward autonomous, tool-using systems able to reason, act and learn within complex technical environments. This evolution opens important opportunities for industrial and application end-users: experts and operators can be supported in time-consuming tasks, interact more naturally with machines, products and digital systems, and improve efficiency, reliability and decision quality.

At the same time, transferring agentic AI into real-world use raises fundamental questions about system architecture, evaluation, robustness, uncertainty management, human oversight, cybersecurity, domainadaptation and deployment under resource constraints.

With this workshop, we aim to narrow the gap between academicresearch and real-world needs. What are the main limitations? Which methods are emerging to address them? Where should we push the scientific and technological SOTA to overcome them?

The discussion will be structuredaround four impulse presentations, each illustrating how CSEM addresses some of thesechallenges throughconcrete applications.

Use case 1 - When Every Byte Matters: Building Conversational AI at the Extreme Edge. Conversational AI is moving from the cloud to increasingly constrained edge devices, enabling private, responsive, and connectivity-independent interactions. Building a complete speech-to-action pipeline on microcontroller-class hardware requires speech recognition, language understanding, local tool execution, and speech generation to operate within strict memory, latency, and hardware constraints. This work examines the practical challenges behind that deployment and the design choices needed to make conversational AI genuinely usable at the extreme edge.

Use case 2 - Medical AI That Knows What It Doesn't Know: Knowledge Graphs, Agents and Human Oversight How can AI systems support healthcare professionals and patients while remaining transparent, traceable, and safe? The CSEM Medical Agentic Solution combines knowledge graphs, agentic AI workflows, uncertainty-aware reasoning, and human oversight to transform trusted medical knowledge into actionable intelligence. Rather than relying solely on large language models, the platform grounds recommendations in structured and traceable knowledge derived from guidelines, protocols, scientific evidence, and organizational expertise. By integrating patient-specific data, medical device data, enterprise systems, and external sources, the solution enables applications ranging from digital medical companions and clinical decision support to medical coding, data structuring, and precision medicine. A key differentiator is its explicit treatment of uncertainty: the system can quantify confidence, identify knowledge gaps, escalate to human experts, or abstain when evidence is insufficient. We present the architecture, discuss the challenges of trustworthy AI in regulated environments, and demonstrate how knowledge graphs, uncertainty quantification, and agentic orchestration can be combined into a reusable foundation for next-generation healthcare solutions.

Use case 3 - Bringing Agentic AI to Industrial Time-Series Data: Challenges and Promises LLM-based agents offer new opportunities for analyzing complex industrial and multimodal data, but deploying them in real-world settings raises important questions around reliability, accuracy, and cost. Using our simulation-based benchmark for root-cause attribution in time-series data, we highlight practical lessons, current limitations, and promising capabilities of agentic AI systems for time-series data in industrialapplications.

Use case 4 - Just Ask: Natural-Language Microscopy Analysis with an AI Agent What if complex microscopy analysis could be performed simply by asking for it? We developed an agent-driven analysis environment that translates natural-language requests into coordinated image segmentation, cell classification, quantitative analysis, and interactive visualization. Behind this seemingly simple interaction lie challenges ranging from reliable tool selection and multi-step workflow management to latency, data heterogeneity, and maintaining consistency across analyses. We showcase our system and discuss the practical solutions that enabled us to turn individual AI models and analytical methods into an integrated scientific workflow.