Theses

In the Internet of Things (IoT) landscape, large networks of sensors collect data that drive crucial insights and decisions. However, the reliability of this data can be compromised due to environmental noise, sensor drift, or hardware inconsistencies. Sensor data fusion—the process of combining data from multiple sensors—can help overcome these challenges by producing a more accurate, consistent, and reliable estimation of the monitored system. The objective of this thesis is to develop a robust mechanism to fuse sensor data, focusing on techniques that validate and enhance data accuracy. By applying sensor fusion techniques, the aim is to design a system that improves data reliability in dynamic IoT environments, creating the basis for smarter, data-driven systems.

Pedagogical Conversational Agents (PCAs) are increasingly utilized in educational settings to facilitate and enhance the learning process. These agents are commonly implemented as text-based chatbots, functioning as conversational interfaces that provide instructional support. PCAs can serve as tutors or motivators, assisting learners in achieving their educational objectives. This thesis aims to integrate anthropomorphic design elements, i.e., features that enhance the agent’s human-like appearance and interaction, into an existing tutoring chatbot. The primary objectives of this research are to conduct a comprehensive review of relevant literature, implement human-like design features, and evaluate their impact on user experience and learning outcomes.

This thesis develops comprehensive assessment mechanisms for evaluating Generative AI-based Internet of Things applications. Evaluating artifacts powered by nondeterministic AI algorithms presents inherent complexity, requiring sophisticated frameworks that address both technical performance and practical effectiveness. The research establishes multi-dimensional evaluation criteria encompassing accuracy, latency, scalability, user satisfaction, and interpretability. These metrics are benchmarked against established standards while adapting recent frameworks for assessing large language models to address the unique temporal and contextual characteristics of IoT data streams. Through systematic review of existing assessment methodologies and analysis of current GenAI-IoT applications, this work proposes novel evaluation approaches tailored to this emerging field. The outcome includes evaluation methodologies, practical guidelines, and tools designed to measure artifact performance in Design Science Research contexts. These assessment mechanisms support iterative refinement cycles and provide structured approaches for validating GenAI-IoT integration effectiveness across diverse deployment scenarios.

This thesis designs and evaluates a conversational interface that enables non-technical users to access and explore IoT sensor data through natural language. It addresses the gap between human-friendly questions and technically complex IoT data platforms (e.g., REST/JSON-based standards such as the SensorThings API), which currently limits direct data access for stakeholders like city administrators or environmental agencies. The proposed solution integrates a Large Language Model with the Model Context Protocol (MCP) to connect natural-language queries to real, structured IoT data sources via tool/API calls. Following a Design Science Research approach with an HCI focus, the work develops a user-centric prototype, demonstrates it in realistic scenarios (e.g., smart city / environmental monitoring), and evaluates it empirically using criteria such as usability, learnability, task accuracy, transparency/trust, and error tolerance. The outcome is a validated artifact plus design principles and recommendations for building reliable LLM-to-API conversational interfaces for structured sensor-data systems.

This thesis investigates how Model Context Protocol (MCP) servers can enable ad-hoc analysis of IoT data stored on SensorThings API-compatible data platforms. While standardized IoT platforms make sensor observations technically accessible, users often still need specialized knowledge to query heterogeneous data streams, combine data sources, and conduct exploratory analyses such as correlations between environmental, spatial, or occupancy-related observations. The work may follow a Design Science Research approach by designing, implementing, and evaluating MCP-based artifacts that connect Large Language Models to IoT data platforms through structured tool/API calls. Building on existing MCP servers for campus-related data sources such as library occupancy, the thesis can extend access to further SensorThings API data sources and support combined analyses across multiple datasets. Depending on the chosen focus, empirical qualitative or quantitative studies may be conducted to gather requirements, assess usability and analytical usefulness, evaluate the reliability of generated analyses, or derive design principles for MCP-based IoT data access and cross-source analysis.

This thesis conceptualizes an interaction model that enables integration between Internet of Things applications and Generative AI technologies. The primary focus is establishing conceptual underpinnings necessary to link IoT sensor data with GenAI-driven processes and interfaces, particularly enabling autonomous and agentic GenAI applications in the IoT domain. The research explores technical foundations including grammar-constrained LLM generation, intermediate prompting approaches using logical programming languages, and emerging standards like Model Context Protocol. These methods address challenges such as ingesting Open Data into standardized platforms, managing heterogeneous data formats, and enabling LLMs to interact with complex distributed systems. The framework incorporates dialog-based interaction models to ensure alignment with natural user query patterns and IoT data interpretation needs.Through systematic analysis of existing interaction models and their limitations, this work proposes a novel framework that facilitates effective communication and collaboration between IoT devices and GenAI systems. The outcome is a conceptual model that bridges technical requirements with user-facing scenarios, establishing foundations for intelligent, context-aware IoT systems enhanced by generative AI capabilities.

This thesis derives comprehensive system requirements for integrating Generative AI technologies with Internet of Things data processing systems. The research draws on multiple sources: literature-based findings, empirical studies of sensor-based use cases in smart city contexts, and prior IoT artifacts from related work. Normative guidelines are established through existing theoretical frameworks, while ethical and societal considerations are integrated throughout the design process, following established practices for multi-user data platforms. The research identifies key challenges, opportunities, and design considerations for developing effective GenAI-based IoT solutions. Through systematic analysis of current technologies, use cases, and theoretical foundations, this work establishes a set of design objectives and requirements aligned with the eDSR methodology. These requirements serve as a foundation for future development in this emerging field, providing guidance for creating efficient and responsible GenAI-IoT integration architectures.

This thesis conducts a comprehensive literature review mapping the landscape of opportunities and challenges when Generative AI interacts with IoT-based systems. While GenAI applications have demonstrated value across various domains, numerous technical and conceptual issues remain unexplored at this intersection. Key challenges include managing the volume and velocity of IoT sensor readings for real-time integration, addressing the unique temporal and contextual characteristics of sensor data that complicate standard GenAI retrieval approaches, and mitigating hallucination phenomena where models generate factually incorrect information. Ethical dimensions around data privacy, fairness, and accountability are examined, particularly in sensitive domains like smart healthcare. The research employs structured literature reviews to identify dominant research streams and gaps in peer-reviewed work, with Smart City applications serving as a primary domain for IoT-based use cases. Qualitative methods including idea-generation workshops and stakeholder interviews map the problem space, while analysis of prototype artifacts and log data provides contextualized, real-world problem evidence. This exploratory, inductive approach follows the eDSR methodology's first echelon.