News

Two new preprints are now available on arXiv, both tackling how machine-learning systems should make decisions when the world is uncertain and some mistakes are far costlier than others.

The first, "Conformal Risk-Averse Decision Making with Optimized Certainty Equivalent Risk Control", asks how an agent should act when it doesn't fully know the state of the system. Rather than fixing a single notion of "cautious," it uses a flexible family of risk measures — the optimized certainty equivalent, which includes mean-variance and conditional value-at-risk as special cases — so a user can dial in how risk-sensitive the decision should be. When the underlying data distribution is unknown, as it usually is in practice, the paper introduces a data-driven calibration procedure that still delivers formal, high-probability guarantees on the chosen risk, with an application to wireless beamforming.

The second, "Risk-Averse Decision Making with Multi-Level Reliability Guarantees", pushes this idea further. Instead of controlling risk at a single level, it provides guarantees at several reliability levels at once, so a user can see in one shot how a decision performs across a whole spectrum from cautious to aggressive. The paper draws an elegant connection to nested prediction sets from conformal prediction, which leads to a tractable optimization procedure, and demonstrates the trade-offs between reliability levels on wireless transmission tasks.

The preprint of "Statistically Valid Hyperparameter Selection: From Tuning to Guarantees" is now available on arXiv. The monograph addresses how hyperparameter selection plays a central role in the development of modern machine learning, and presents a unified framework for obtaining statistical guarantees on the selected hyperparameters.

I have started a Research Associate position at Northeastern University London within the Institute for Intelligent Networked Systems (INSI), supported by European Research Council funding. Excited to continue work on trustworthy AI and statistical inference in this new environment.

Collaborated with Oriel College, University of Oxford on a video discussing autonomous vehicle technology, addressing public questions regarding safety, regulation, and deployment ahead of the London launch of self-driving cars.

Delivered a tutorial presentation titled "From Tuning to Guarantees: Statistically Valid Hyperparameter Selection" at NeurIPS 2025. The tutorial covered the theoretical foundations of conformal and hypothesis-testing-based approaches to hyperparameter optimization with formal statistical guarantees.

Excited to announce a 3-hour tutorial at NeurIPS 2025 on moving beyond empirical methods toward statistically guaranteed hyperparameter selection. The tutorial bridges classical statistical testing with modern machine learning practice.

Competed in the University of Oxford Three Minute Thesis finals, earning the runner-up prize among eight finalists. The presentation distilled the core ideas behind statistical guarantees for machine learning into an accessible three-minute talk for a general audience.