Postdoctoral Fellow at Mila – Quebec AI Institute, co-supervised by Yoshua Bengio and David Rolnick.

During my post-doc I developed AI systems to reduce carbon emissions from buildings—which account for ~37% of global CO₂ emissions. My research focuses on graph-structured world models and reinforcement learning, with an emphasis on systems that generalize (for example, across different building types, network layouts, and environments). Previously I studied causal inference and Bayesian methods for decision-making.

My work sits at the intersection of model-based RL, graph neural networks, and physics:

• G Dreamer — A world model architecture for variable-size, heterogeneous graph environments that learns structural relationships governing thermal dynamics, enabling zero-shot transfer across buildings with different topologies
• HOT Dataset — A dataset of ~150,000 simulated controllable buildings and gym for transfer learning research
• ARC playground — A testbed for exploring learning concepts as invariants / constraints instead of conventional labels

I'm particularly interested in how we can learn generalizable physics from simulations.

• HVAC Spice — UrbanAI @ NeurIPS 2025
• Graph Dreamer: Temporal Graph World Models for Sample-Efficient and Generalisable RL — WiML @ NeurIPS 2025 [OpenReview]
• A HOT Dataset: 150,000 Buildings for HVAC Operations Transfer Research — BuildSys 2025 [HuggingFace]
• Challenges and Opportunities of IoT in Occupant-Centric Building Operations — Current Opinion in Environmental Sustainability, 2023 [Paper]

📚 Full publication list on Google Scholar

Graph Neural Networks  •  Model-Based RL  •  EnergyPlus  •  PyTorch Geometric  •  Spatiotemporal Modeling

Most of my recent code lives in private repositories (pre-publication research), but I'm working on releasing components as papers are published.

Two-time Olympian in modern pentathlon (London 2012, Rio 2016). Currently based in Nice, France. Board member at Racing to Zero, a nonprofit focused on sustainability in sport.