Product marketing engineer at Langfuse, teaching developers how to inspect and evaluate AI applications.[1]
Start with Yes, you can copy our eval setup →
Education
Links
Team and careers Speaker profile
X · annabellschfr 
GitHub · annabellscha Product Marketing Engineer · Langfuse
Career
At Langfuse, now part of ClickHouse, Schäfer combines technical writing and developer education with practical examples of application monitoring. She jointly introduced Langfuse Academy with Lotte Verheyden, connecting traces, datasets and experiments so teams can investigate actual agent failures and test changes before deployment.[4]
Contributions
Turning conversations into tests
Her documentation-chatbot evaluation guide starts with traces of retrieval, tools and generated answers. Narrow detectors flag conversations for review, while a general sample helps find failures the detectors miss. Human reviewers label outcomes and explain what went wrong; those examples then become evaluation datasets. This separates a retrieval failure from a poor answer based on good evidence, which require different fixes.[3]
Finding patterns across agent traces
Her process-mining cookbook provides a Jupyter notebook and sample Langfuse traces for inspecting recurring sequences of agent activity. The accompanying flow visualization and exported analysis give developers a way to examine repeated paths across runs, beyond reading one conversation at a time.[2]
Tools, writing & talks
Writing, research & talks
Start here
Yes, you can copy our eval setuplangfuse.comAn authored walkthrough of production traces, monitoring signals, human review and datasets for a documentation chatbot.
Read the chatbot guide with an application failure in mind. It explains what to trace and review before choosing scores; following the worked setup assumes familiarity with an LLM application and Langfuse.
Author[3]- Stop Burning Tokens: Why self-improvement needs domain expertise first - Annabell Schäfer, Langfuseyoutube.com
Langfuse ran a minimal self-optimization loop on paper classification against ground truth and got a 15% accuracy gain — then unpacks what made it work: a narrow task, a real target function, explicit stopping criteria, and a clear handoff point between domain experts and automation.
Catalog workOriginal source
Tools
- Agent Process Mining Cookbookgithub.com
A notebook and example traces for applying process-mining techniques to Langfuse agent data.
Author and repository owner[2]