At Leiden University, I compared semantic retrieval with tool- and graph-guided coding agents for repository question answering. My research found that simple retrieval was more efficient, while graph guidance doubled the agents’ rate of fully correct answers.
I also built an LLM-as-a-Judge framework for scalable, reproducible code-understanding evaluation. In other projects, I designed Codenames agents and built a convolutional neural network to read analog clocks from images.

