Projects
Publications & Preprints
AI for Science
MA-HGRL: Memory-Augmented Hierarchical Graph Reinforcement Learning for BOM Optimization
A hierarchical graph RL agent with an external memory module for optimizing bills of materials. Extended archival version in preparation for TMLR.
BE-MDP: Budgeted Exploration in Volatile MDPs
Offline-to-online reinforcement learning with an adaptive Lagrangian for exploration under a hard budget, in environments whose dynamics shift over time. Under submission to AAAI 2027.
Physical vs. Informational Leakage in Circular Supply Chains
A formal decomposition separating material that truly leaves a circular system from material that only disappears from its records, with estimation methods over ERP transaction data.
About
I'm a PhD researcher at CentraleSupélec (LGI, Université Paris-Saclay), working with GE HealthCare on AI-driven digital twins for the circular supply chain of MRI and CT spare parts. The core question: when a part disappears from a global system, did it physically leave, or did the data just lose track of it?
My toolkit sits at the intersection of reinforcement learning, graph machine learning, and probabilistic modeling, applied to messy, large-scale industrial data. On the side, I care a lot about language models that work natively in Arabic and other underserved languages, which is where my NLP projects live.
Before the PhD: a double engineering degree from École Centrale Casablanca and CentraleSupélec (AI specialization), with stops at SLB, Carrefour, and Illuin Technology along the way.