Safety filters that let a learning controller improve performance while keeping a checkable safety guarantee, and a bound on how much safety must be given up when the safe data hide what could go wrong. Validated on insulin dosing, a chemical reactor and CartPole. Paper in preparation.
I'm Guglielmo Cimolai, an aerospace engineer working where physical simulation meets AI. I co-founded coverd.ai and have just completed an MSc in AI at Imperial College London. This is my website, where I share my research, the projects I take on, and the occasional thought.
I'm an aerospace engineer based in London, working where physical simulation meets AI. I've just completed an MSc in Artificial Intelligence Applications and Innovation at Imperial College London, with a thesis on Reinforcement Learning for safe control of critical systems.
I'm co-founder of coverd.ai, where I lead AI and research: we build voice AI that runs structured interviews and turns them into knowledge.
Before that I spent three years at Airbus Helicopters, as a consultant with Digital Product Simulation, building a Simulation Process and Data Management platform used regularly by ~60 engineers. Earlier, my MSc at Cranfield University led to four journal papers in structural and computational mechanics.
Click-to-track video system (YOLO, BoT-SORT, SAM 2) with game-disjoint, traceable HOTA evaluation.
An AI voice interviewer that turns structured conversations into a knowledge graph, linking each claim to its evidence in the interview. I designed the architecture.
Open benchmark testing whether LLMs keep rules embedded in realistic policies under user pressure: 11 frontier models, 129 rules, LLM judges calibrated against blind human labels (κ 0.74–0.79). Polite multi-turn pressure raised violations ~5×. The v0.7 technical report and dataset are public (DOI 10.57967/hf/9330); ported to UK AISI Inspect Evals (under review).

