SA/BA im Bereich Statistical Machine Learning / Deep Learning

This project presents a proof-of-concept conversational coach that supports apprentices in producing structured learning journal entries and linking them to profession-specific training plans. The work addresses common shortcomings in practice, where documentation is often created too infrequently and lacks structure, linguistic quality, and reflective depth due to the difficulty of starting from a blank page.

The prototype initiates reflection from minimal, unstructured input and guides apprentices through targeted follow-up questions. It incorporates the relevant training plan into the prompt and can transform the completed conversation into a journal entry while suggesting suitable competences for linkage. To enable continuity across a traineeship, past journal entries are indexed in a vector store and retrieved per message via similarity search. The concept was evaluated through an automated test setup that emulates apprentice–coach conversations. It measures semantic alignment between situation descriptions and generated journals using embedding cosine similarity, and evaluates competence suggestion quality by computing precision, recall and F1 score against a set of expected competences. A qualitative user test with one apprentice complemented the automated evaluation and provided early usability feedback.

Results indicate that the conversational approach lowers the barrier to starting documentation and can generate complete, well-structured journal entries, while competence suggestions remain more variable and require human oversight. Future work includes prompt refinement, temporal anchoring through explicit inclusion of the previous journal entries, broader evaluation with more apprentices, and assessing integration strategies for deployment within Salto.