I'm Henri! I'm passionate about AI safety, particularly AI control, and reducing AI risks. To that end, I've been working as a Research Engineer at EquiStamp on LinuxArena, a new AI control setting for live production software environments, in collaboration with Redwood Research. I've also baselined for METR through EquiStamp. I hold a B.Sc. in Statistics and Computer Science from McGill.
Recent projects include: LinuxArena, Control Tower, Untrusted Editing protocols & exploring feedback exploitation in control protocols, bisampling protocols, and AI control scaling laws.
I've worked on tools like AlignmentSearch and Stampy Chat to help others learn about alignment. During my undergrad, I did a research project on sim-to-real transfer for robotic locomotion under Prof. Hsiu-Chin Lin.
I organize ACX Montreal meetups
and previously helped run EA McGill and founded AI Alignment McGill. In my free time, I enjoy chess,
arm wrestlingI stopped arm wrestling when I broke my arm at Manifest. :P, and forecasting.
Publications
Projects
AI control evaluation library powering LinuxArena
AI control setting measuring covert sabotage by AI agents in live production software environments
A real-time multiplayer infinite minesweeper game. Players explore an unbounded world together, competing on a global leaderboard.
Developed feedback-based AI control protocols
Open Low Cost Humanoid
Developing accessible, open-source humanoid robot with PPO-based locomotion for sim2sim and sim2real transfer
Lifelogging
High-performance Rust-based continuous audio recording with SIMD optimization
A benchmark evaluating LLM friendliness — scoring models on sycophancy resistance, conversational warmth, and genuine personality.
A retrieval-augmented generation platform that helps users explore AI safety research through conversational interface
I've also worked on some less serious projects: PressBench (benchmarking AI bench-press self-assessment) and an n-dimensional chess calculator.