Experience & education

• Introduced Magentic's first agent and lead the ongoing development of its architecture, validating design choices through experimentation against eval suites built from real procurement data
• Continuously improving system reliability through an iterative loop combining hillclimbing, feedback integration, and model routing as new data and edge cases emerge
• Embedding human-in-the-loop design throughout, balancing agent autonomy with oversight to match the risk profile of procurement and supply chain stakeholders

• Led the research team developing agentic red-team evaluations for LLMs, targeting custom safety specifications in conversational and tool-using agents
• Designed automated pipelines that 3x'd the coverage of red-teaming techniques in 6 months, increasing behavioral diversity across evaluation scenarios
• Contributed to core research efforts, including a first-author paper on the robustness of LLM safety judges and a medical benchmarking hallucination paper

Worked on efficient methods to perform zero-shot and weakly-supervised referring image segmentation (i.e., segmenting an object in an image that is referred to in a natural language sentence), achieving new state-of-the-art performance in the field.

Extended the certified robustness technique of randomized smoothing from isotropic ℓp balls to anisotropic certificates through a simplified Lipschitz analysis-based framework.

Thesis: "Towards Trustworthy Machine Learning Models in Vision, Physics, and Language Applications"
Relevant courses: Data Estimation & Inference, Machine Learning, Reinforcement Learning, Discriminative & Deep Learning for Big Data
Supervisors: Prof. Philip H.S. Torr, Dr. Adel Bibi and Dr. Pawan Kumar (Google DeepMind)

• Led the development of safe and scalable optimization-based motion planning algorithms, working in a team with research scientists and software engineers
• Published and presented research work developed at top tier conferences and journals within the robotics community, as well as to non-technical audiences
• Wrote and reviewed research and development code, ensuring CI with other tools within the company

Relevant courses: Computational Game Theory, Probabilistic Model Checking
Dissertation: "To Err is Human: Designing Correct-by-Construction Driver Assistance Systems using Cognitive Modelling"
Supervisors: Dr. Morteza Lahijanian and Prof. Marta Kwiatkowska

Relevant courses: Applied Machine Learning, Image and Video Processing, Lab in Image and Signal Processing
Awarded a monthly scholarship under the Swiss-European Mobility Programme

Relevant courses: Algorithms and Data Structures, Signals and Systems, Computational Mathematics, Automatic Control
Top 2% of the class
Academic Excellency Award every year and for the end of the BSc (3 years)