
Silent Failures in Agentic DS
Evaluating data science agents beyond accuracy: most agent pipelines hide at least one methodological failure, such as data leakage
Some representative projects I have worked on. The full publication list is on the about page and on Google Scholar.

Evaluating data science agents beyond accuracy: most agent pipelines hide at least one methodological failure, such as data leakage
Systematically identifying distribution shifts and their impacts in NER datasets

Efficient data wrangling with LLMs by generating code instead of prompting row by row

Knowledge-centric prompt composition for knowledge base construction, 2nd place on Track 2
Knowledge graph construction from conversational data, my PhD project