This thesis investigates potential bias in AI-based hiring systems, focusing on gender and nationality discrimination. In particular, it audits a screening pipeline built on Sentence-BERT (SBERT), which compares CVs and job descriptions using a semantic similarity score. As these tools are increasingly used to support shortlisting decisions, they may unintentionally reflect or reinforce existing unfair patterns, leading to different outcomes for different groups. The aim of this work is to propose a clear and reproducible audit approach to evaluate fairness in a hiring-like scenario. To do this, we set up a controlled experiment in which the synthetic candidate profiles are written consistently, so that the only meaningful changes are those related to gender and nationality.