events
Building AI Evaluations You Can Actually Trust
An online episode from the Berlin DataTalksClub Group in which Tejas Kumar, an AI engineer at IBM, looks at why common ways of evaluating language models give misleading results and how engineering…

Auf Englisch
- Starts
- 2026-10-13 10:30 UTC
- Organizer
- Berlin DataTalksClub Group
- Location
- Online event
An online episode from the Berlin DataTalksClub Group in which Tejas Kumar, an AI engineer at IBM, looks at why common ways of evaluating language models give misleading results and how engineering teams can make them more dependable. He covers calibrating AI judges against human labels and strict rubrics, countering length and position bias, replacing a single fuzzy score with instant deterministic checks, and assessing retrieval-augmented generation stage by stage to separate retrieval failures from generation failures. The link between agent harnesses and successful evaluations is also discussed.
Event website