Research
We propose a hierarchical learning and planning framework for large-scale power-grid
control. A heterogeneous GNN first identifies the most relevant subgraph to act on,
reducing the action and observation space. A tree-search procedure is then applied
locally over this region to evaluate topological actions, showing that combining
structure-aware graph representations with search improves efficiency and
generalization across different grid topologies.
We introduce a multi-agent, multi-stage pipeline for controllable molecule generation,
where specialized agents collaborate to satisfy multiple structural and chemical
constraints. The system progressively refines molecules across stages, achieving both
diversity and precise satisfaction of property targets, offering a scalable framework
for realistic multi-objective molecular design.
Projects
These are some relevant projects I worked on during my master's and undergraduate degrees.
Bayesian Graph Neural Networks for the Optimal Power Flow Problem
Mila · 2024
Explored uncertainty-aware graph learning methods for power system optimization using
Bayesian GNNs on the Optimal Power Flow (OPF) problem.
-
Developed a Bayesian GNN model combining approximate inference techniques such as
variational inference, Monte Carlo sampling, Monte Carlo dropout, and ensemble methods
to capture predictive uncertainty.
-
Implemented a baseline following the CANOS paper (Piloto et al., 2024) using the
OPFDataset to benchmark performance on OPF.
-
Applied constraint-aware learning with Lagrangian relaxation and dual methods in the
loss function to enforce physical and engineering feasibility.
-
Demonstrated that the Bayesian approach improves generalization, provides calibrated
uncertainty estimates, and enhances robustness for high-stakes power grid decisions.
Autonomous Driving – Duckietown
Université de Montréal · Duckietown Class · 2024
Applied simplified inverse reinforcement learning (IRL) and DrQv2 for autonomous driving
control in the Duckietown environment.
-
Implemented a simplified IRL approach combining a predefined reward function with
learned residuals to preserve interpretability while improving performance.
-
Used expert trajectories to fine-tune the reward function so the agent learns
realistic driving constraints.
-
Showed that the IRL agent learned subtle behaviours such as staying in the right lane,
while a pure RL agent tended to drive on the left.
Exploration Drones – Web-Controlled Multi-Drone System
Polytechnique Montréal · Jan 2023 – May 2023 · Python · TypeScript · C++
Built an interactive system to control exploration drones and visualize their mapping of
unknown environments.
-
Developed a Vue-based web application to control drones and display their live
exploration map.
-
Implemented the exploration algorithm on the embedded system to explore and map
unknown terrain while avoiding obstacles in multi-drone scenarios.
Package Delivery Drone System for Montréal
Data Structures Project · Polytechnique Montréal · May 2020 – Aug 2020 · Java
Designed algorithms to reduce energy consumption in a fleet of delivery drones.
-
Assigned packages to drones based on weight, size, availability, and route cost using
Dijkstra’s shortest path algorithm.
- Improved overall fleet efficiency by optimizing routing and load balancing.
Embedded Robot for Covid Distancing
Embedded Systems Project · Polytechnique Montréal · Jan 2020 – May 2020 · C++
Developed an embedded robot that maintains preventive distances from people in a crowd.
-
Programmed a finite state machine defining the robot’s reaction to six different
proximity scenarios.
-
Improved development efficiency by scripting automated deployment of the embedded
code.
