• Trustworthy AI in hiring and recruitment: LLM-based systems are increasingly deployed in high-stakes domains such as hiring pipelines, where their decisions directly affect people. This project studies how trustworthy these systems are in such settings with a focus on their fairness and security. It is funded by and carried out in collaboration with Amazon.

• Bias and fairness in large language models (LLMs): bias in an LLM is not spread evenly across the network but concentrated in particular components, so the work first locates which attention heads and feed-forward layers carry it, then fine-tunes only those components against a fairness-aware training objective (ongoing).
• Privacy–fairness tensions in distributed ML: differential privacy often widens performance gaps across demographic groups, creating fairness issues. To address this, we designed a group-aware model architecture that reduces performance gaps across groups under differential privacy, without sharing group information. Publication under review.

Research and development across two EU-funded projects:

• Enterprise software development (ERP): built new Odoo modules end to end (Python backend, JavaScript front end), supporting finance and supply-chain operations for business clients.
• Systems integration: connected those modules to clients' existing finance and supply-chain systems, keeping data and processes consistent across both.

• Anomaly detection in information networks: developed graph-based methods that flag suspicious nodes by tracking their activity and how it changes over time, an approach applicable to fraud detection, security monitoring, and abuse detection in social networks.