Bachelor Thesis

Fleeting verbal agreements in agile software development often facilitate a gradual loss of knowledge, frequently termed "knowledge vaporization", as manual documentation in practice often remains incomplete. This bachelor thesis addresses this gap through the design, implementation, and evaluation of a system for the automated extraction of Scrum artifacts (Action Items, Decisions, Impediments, Risks) from meeting transcripts.

Following the Design Science Research methodology, a modular pipeline was developed that integrates speech-to-text technologies with the reasoning capabilities of Large Language Models (LLMs). The study systematically contrasts two prompt engineering strategies: direct extraction via Zero-Shot and the analytical Chain-of-Thought approach.

The evaluation, based on a hybrid dataset of real-world and synthetic meetings, demonstrates a significant divergence in performance: while Zero-Shot operates efficiently in standardized contexts, the Chain-of-Thought approach ensures higher precision in complex, unstructured situations. It identifies implicit causalities and sarcasm while reducing hallucinations during domain shifts. The results imply that LLMs, acting as assistance systems, can substantially improve documentation quality, provided that the prompting strategy is adapted to the complexity of the communication.