More about me

For a longer version including trivia and pics, see Even more about me.

I am originally from Trieste, Italy, but I spent the last 10 years hopping around the world for my studies. I graduated from an undergrad CS program in Québec, Canada, before returning to my hometown for an MSc in Data Science at the University of Trieste and SISSA. My master’s coincided with the first pre-trained Transformers (BERT, GPT-2), so I started working in NLP relatively early — first as a research assistant in the ItaliaNLP lab of CNR-ILC,¹ and then as a research scientist working on generative models for text and vision at Aindo.

In 2021, I moved to the University of Groningen to start a PhD with Arianna Bisazza, Malvina Nissim and Grzegorz Chrupała as part of the GroNLP group and the InDeep consortium. My PhD project focused on actionable interpretability for machine translation, and in 2022 I spent three nice months in NYC for an internship at Amazon Translate.

Over the course of my PhD, I became increasingly interested in building open-source tools and interfaces to democratize access to interpretability methods by improving their usability, efficiency and scalability. This led me to develop Inseq, a library for conducting attributional analyses of LLMs, and to contribute to the development of other interpretability tools such as NNsight and Interpreto.

After finishing my PhD in 2025, I joined the Bau Lab of Northeastern University to work on scalable interpretability methods and interfaces for the National Deep Inference Fabric (NDIF) team. There, I helped design the NDIF ecosystem and run the Aletheia’s Quest deception detection competition.

Now

(Last updated: August 2026)

I am still a postdoctoral researcher at the Bau Lab. These days, my work focuses on scalable methods and infrastructure for analyzing the internal beliefs, plans and goals of AI agents – an area that has become increasingly important in light of recent evidence of evaluation awareness, emergent misalignment and other worrying phenomena in frontier systems.

In late 2026, my friend and colleague Mario Giulianelli and I also co-founded Parallax, a London-based non-profit research organisation aiming to make white-box auditing a standard component of frontier AI evaluations. Parallax is still taking shape, but I foresee working on it full-time in the near future.