Built an AI coaching platform combining long-term user memory, RAG-based personalization, structured coaching sessions, and challenge-cycle orchestration. Implemented a 6-layer memory pipeline from raw conversation data into atomic claims and rolled-up summaries, using PostgreSQL, pgvector, and embeddings. Developed LangGraph/LangChain orchestration with async memory processing, structured Pydantic outputs, and cost-aware model routing, including human-in-the-loop workflows with confirmation gates, token-budgeted context assembly, and audit trails. Created LLM evaluation workflows using golden datasets, RAG quality metrics, LLM-as-judge checks, and Langfuse traces for quality/cost monitoring.

Teaching assistant of Mathematics at University of Bern

Supported a German-language Discrete Mathematics course (84 students), gained Swiss work experience in a high-standard academic environment, and demonstrated adaptability.