AI / ML Researcher
I work across the full model lifecycle, from large-scale self-supervised pre-training of multi-modal and world-model backbones, to post-training that couples learned representations to reinforcement-learning agents.
I am an AI/ML Researcher at King AI Labs (Microsoft Gaming), where I build predictive world models on heterogeneous game signals and connect them to in-game decision-making. My work is grounded in the theoretical foundations, generalization, and safety of modern deep learning, and validated at scale in large gaming environments.
I previously completed my PhD in the School of Electrical Engineering and Computer Science at the KTH Royal Institute of Technology, supervised by Professors Michalis Vazirgiannis and Henrik Boström, and funded by the Wallenberg AI, Autonomous Systems and Software Program (WASP). My dissertation, On the Adversarial Robustness of Graph Neural Networks, developed theoretically grounded attacks and defenses, work that now informs how I think about the safety of pre-trained systems.
I also had the pleasure of spending a summer at the Flatiron Institute (Simons Foundation) with the Polymathic AI / Foundation Models for Science initiative, working with Leopoldo Sarra and Siavash Golkar on extending Joint-Embedding Predictive Architectures (JEPA) to time series (TS-JEPA).
Earlier, I earned an MSc in Applied Mathematics from École Polytechnique and an engineering degree from EMINES at Mohammed VI Polytechnic University (UM6P), Morocco.
Research
One thread runs through my work: how representations are learned at scale, connected to agents, and made reliable.
Self-Supervised & Multi-modal Pre-training
Predictive world models and JEPA-style objectives; self-supervised representation learning across modalities and temporal graphs, with a focus on representation collapse and normalization.
World Models & RL
Context-conditioned world models that bridge representation pre-training and policy learning by coupling pre-trained backbones to reinforcement-learning agents for in-game decision-making (model-based RL).
Generalization & Robustness
Theoretical foundations of Transformer-based models, including expressivity of pooling and preprocessing effects, alongside the adversarial robustness and safety of deep learning systems.
News
- 2026 · "Understanding Early Collapse in Predictive World-Model Pretraining" accepted to the World Models Workshop @ ICLR 2026.
- 2026 · "Be Wary of Your Time Series Preprocessing" accepted to the AI4TS Workshop @ AAAI 2026.
- Sep 2025 · "Virtual Nodes Go Temporal" accepted for an Oral presentation at LOG 2025!
- Sep 2025 · Two papers accepted at NeurIPS 2025.
- May 2025 · Talk on GNN robustness at the Metis Spring School, Rabat, Morocco.
- Feb 2025 · Survey "Expressivity of Representation Learning on Continuous-Time Dynamic Graphs" accepted to TMLR with a Survey Certification.
- Sep 2024 · "Joint Embeddings Go Temporal" accepted to the Time Series in the Age of Large Models Workshop @ NeurIPS 2024.
- Sep 2024 · "If You Want to Be Robust, Be Wary of Initialization" accepted to NeurIPS 2024.
Selected Publications
-
Learning representations of Continuous-Time Dynamic Graphs (CTDGs) has gained attention for modeling real-world phenomena. We conduct a theoretical analysis of Virtual Nodes (VNs) in CTDG-based models via a notion of information flow (how information propagates after an event), highlighting limitations of existing approaches and how VNs address them. We propose k-TVNs, a framework of fully connected VNs each representing a community, and show both theoretically and empirically that VNs substantially improve CTDG models.
-
To bring JEPA to time series, we introduce Time Series JEPA (TS-JEPA), performing self-supervised learning in latent space rather than reconstructing inputs. We validate TS-JEPA on classification and forecasting, matching or surpassing state-of-the-art baselines and showing strong balance across diverse tasks, laying groundwork for time-series foundation models based on Joint Embedding.
-
A comprehensive review of Graph Representation Learning on CTDGs with a focus on Self-Supervised Representation Learning. We introduce an Information-Flow framework analyzing the expressivity of CTDG models by quantifying their ability to propagate and encode temporal and structural information, and categorize existing methods by suitability across long-range, bipartite, and community-based graphs. Awarded a Survey Certification.
-
We introduce a theoretical framework characterizing the expressivity of Transformer-based models under widely used pooling methods, deriving closed-form bounds on representational capacity and the ability to distinguish similar inputs, across attention variants. Empirically, across vision, NLP, and time-series, we reveal consistent trends in how pooling affects accuracy, sensitivity, and optimization, positioning pooling as a key architectural component beyond attention alone.
-
We analyze the overlooked role of pooling in the adversarial robustness of graph classification, deriving upper bounds on the adversarial risk of standard flat pooling methods. Motivated by these insights, we propose Robust Singular Pooling (RS-Pool), which leverages the dominant singular vector of the node-embedding matrix. RS-Pool is model-agnostic, implementable via power iteration, and empirically more robust under state-of-the-art attacks while maintaining competitive clean accuracy.
-
We study an under-explored dimension of robustness: weight initialization and training epochs. We introduce a theoretical framework relating initialization strategies to resilience against adversarial perturbations, revealing a direct relationship between initial weights, training length, and vulnerability, with a general upper bound applicable to deep neural networks. Experiments show appropriate initialization improves robustness by up to 50% over alternatives.
-
We define expected robustness for attributed graphs and derive an upper bound for GCNs and GINs under node-feature attacks. Connecting robustness to the orthonormality of weight matrices, we propose GCORN, an attack-independent, more robust GCN variant, plus a probabilistic method to estimate expected robustness. GCORN outperforms available defense methods across real-world datasets.
-
We introduce NoisyGNN, a defense that injects noise into the model architecture, with a theoretical connection between noise injection and improved GNN robustness. Extensive node-classification experiments on GCN and GIN show superior or comparable defense to existing methods while minimizing added time complexity. NoisyGNN is model-agnostic and composes with other defenses for further gains.
* denotes equal contribution.
Experience
AI/ML Researcher · King AI Labs (Microsoft Gaming)
- Multi-modal pre-training & representation. Built self-supervised, multi-modal pre-training that learns unified representations from heterogeneous game signals, improving downstream player embeddings and recommendation.
- World models, from pre-training to RL. Pre-trained predictive world models on multi-modal game data, extending temporal JEPA and mitigating early representation collapse, then coupled them to RL agents via context-conditioned world models (model-based RL).
- Architecture & training foundations. Grounded these systems in theory, particularly pooling in Transformers and normalization/preprocessing effects, to improve pre-training stability and quality.
Research Intern · Flatiron Institute (Simons Foundation)
- Polymathic AI / Foundation Models for Science initiative.
- Extended JEPA from vision to sequential data, proposing TS-JEPA: a self-supervised framework that learns time-series representations by predicting in latent space rather than reconstructing inputs.
Research Intern · BNP Paribas
- Developed counterfactual-explanation methods for black-box models: the minimal input changes that flip a decision.
- Made financial decision systems interpretable and auditable for regulatory compliance.
Research Scholar · University of Louisville
- Designed an end-to-end deep-learning CAD pipeline for lung-nodule detection from 3D CT scans.
- Reached a top-tier ≈94% (±0.6) accuracy on the LUNA challenge benchmark.
Technical Expertise
Pre-training
Self-supervised and multi-modal pre-training of LLM, world-model, and time-series backbones (JEPA, DINO), including objective design and representation-collapse / normalization diagnostics.
Post-training
Connecting pre-trained world models to reinforcement-learning agents (model-based / context-conditioned RL) and fine-tuning foundation models for real-world applications.
Scale & Infrastructure
Large-scale distributed training (DDP) and end-to-end experimentation on HPC clusters (A100 / H100), from data ingestion and experiment tracking to reproducible evaluation.
- Python
- PyTorch
- DDP
- A100 / H100
- JEPA
- Model-based RL
Teaching & Service
Teaching
- Graph Curvature as a Lens on Adversarial Robustness in GNNs · LOGML Summer School, Imperial College London.
- Introduction to LLMs & Deep Learning on Graphs · École Polytechnique, Paris.
- Deep Learning for Time Series, NLP and Graphs · École Polytechnique Executive Education, Paris.
Selected Talks
- Representation Learning and World Models for Games · Imperial College London 2026.
- Virtual Nodes Go Temporal · Temporal Graph Reading Group, 2026.
- From Bounds to Defenses: A Comprehensive Look at GNN Robustness · Metis Spring School, 2025.
- Theoretically Upper-Bounding the Expected Adversarial Robustness of GNNs · Collective ML, 2024. [Slides]
- Adversarial Robustness of GNNs · MoroccoAI webinar. [Slides | Recording]
Reviewing
NeurIPS (Top Reviewer, 2025), ICLR, ICML (Top Reviewer, 2026), AAAI, and TMLR.
Awards & Honors
- WASP Doctoral Scholarship · Knut and Alice Wallenberg Foundation, 2021.
- OCP Full Excellence merit scholarship for outstanding entrance-examination results, 2014.