DBIS

With rapid advances in artificial intelligence, especially large language models (e.g., OpenAI GPT, Anthropic Claude, Google Gemini), AI has demonstrated great potential in natural language processing, code generation, and automated testing. Programming, traditionally a highly specialized and creative task, is increasingly supported and partially automated by AI. Tools such as GitHub Copilot and OpenAI Codex already generate code snippets based on natural language prompts, boosting developer productivity.

However, current AI code generation systems face challenges such as limited context understanding, inconsistent code quality, lack of adaptation to complex software engineering workflows, and limited capacity for continuous interaction and collaboration. Therefore, building an intelligent coding agent system capable of autonomous planning, requirement comprehension, code writing, debugging, and optimization has become a cutting-edge research topic in academia and industry.

As software engineering continues to evolve, agile frameworks have become the main paradigm for project management and development. Among these, Scrum is widely adopted by numerous software organizations due to its iterative and incremental development approach, emphasis on team collaboration, and rapid feedback cycles. With the ongoing advancements in artificial intelligence technologies, particularly the recent breakthroughs in large language models, there is growing interest in integrating AI throughout the entire lifecycle to improve efficiency and quality further.
 

While recent advancements in natural language processing have been largely driven by increasingly powerful large language models (LLMs), the role of data quality in fine-tuning these models remains underexplored. This thesis addresses the often-overlooked but critical aspect of data-centric AI by investigating how different types and levels of data degradation affect the performance of fine-tuned LLMs on tasks such as summarization and question answering. Unlike model-centric approaches that focus on architectural improvements, this work systematically introduces controlled degradations, such as spelling errors, grammatical mistakes, and semantic noise, into high-quality datasets. The goal is to quantify how these degradations impact downstream performance and identify which types of data noise are most detrimental. The results aim to inform best practices in dataset curation and reinforce the importance of data quality in building robust, task-specific LLM applications.

Technical support is an essential aspect of various industries, e.g., to provide help with maintaining machinery and IT systems. However, diagnosing error messages and faults in complex technologies can be a time-consuming and challenging task. The maintainer has to search through the long documentation booklets for the technology in order to find a solution or wait for an appointment so that the manufacturer sends an expert. A combination of mixed reality (MR) agents and large language models (LLMs) offer potential to provide quick and interactive diagnosis suggestions while providing visual help. For instance, the MR agent could be placed next to the faulty machine and embody the LLM. The technical person could talk to the agent which, in turn, forwards the inputs to an LLM. Through techniques such as retrieval augmented generation (RAG), the LLM could match the described problems with suitable excerpts of the documentation. Thus, it can quickly identify relevant sections in a long document and provide help based on the knowledge that is provided in the documentation. The answer could then be output by the MR agent and it could be combined with pre-defined actions by the MR agent, e.g., to point to specific points on the machine or to demonstrate a process on a virtual replica.

The innovative integration of Mixed Reality and Large Language Models can lead to highly interactive instructional MR agents. Utilized as automated instructors, these MR agents have the potential to significantly enhance traditional instruction manuals by providing visual guidance. For instance, they can illustrate the next required actions in practical tasks such as tightening screws in machine maintenance. With LLMs, the interactivity of the MR agents can further be enhanced by enabling users to engage in a dialogue with the MR agents, posing questions and receiving real-time responses. Here, a challenge lies in providing a spatial understanding to the LLM so that it can refer to elements in the MR space.

The increasing connectivity of Industrial Control Systems (ICS) has elevated the need for robust cybersecurity measures. However, evaluating the effectiveness of Intrusion Detection Systems (IDS) in ICS environments remains fragmented and inconsistent. This thesis addresses this challenge by developing a systematic, modular benchmarking environment that enables reproducible and standardized evaluation of machine learning-based IDS across diverse datasets. By introducing a unified data format and structured evaluation protocol, the work aims to enhance the comparability, transparency, and practical relevance of ICS security research.

The goal of this thesis is to design, implement, and evaluate a sensor-based maintenance bot that uses Large Language Models (LLMs) to support predictive maintenance and decision-making. The bot should be capable of retrieving, analyzing, and reasoning over time series sensor data as well as unstructured maintenance-related documentation (e.g., technical manuals, incident reports). The result is a unified system that assists technicians and engineers in diagnosing issues, suggesting preventive actions, and retrieving relevant information in natural language.

This work is conducted in collaboration with the Fraunhofer Institute for Production Technology (IPT) as part of the research initiative “Generative AI for Production and Business Operations“, aiming to explore practical applications of generative models in manufacturing and industrial operations.