Sprecher
Beschreibung
We apply AI to the education domain, aiming to prepare high-quality digital materials for students—specifically transforming lectures from a hadron physics course into reliable resources. The core challenge is to combine several streams of information (speech, formulas, figures) into a coherent product, which cannot be solved by a single-prompt approach. Our pipeline employs embedding-based semantic chunking to group related content within strict context limits, while preserving the lecturer’s phrasing, explanatory style, and technical detail. Mathematical expressions are recovered as LaTeX with contextual accuracy, and figures are automatically integrated into the narrative. The system is implemented in a hybrid stack (Python, Deepseek API models, ollama interface for embeddings), enabling reproducible and extensible workflows. Beyond classroom use, the approach points toward future applications such as automated preparation of conference proceedings in the STEM domain.