Research
My research sits at the intersection of applied mathematics and deep learning, with a focus on making AI systems safe and reliable in the presence of adversaries. The guiding question: how can we prove, and not just hope, that learning systems remain robust?
Research Interests
Adversarially Robust Distributed Learning
Distributed training scales learning across many workers, but a single malicious worker can corrupt the model through crafted gradients. I study Byzantine-resilient aggregation rules (Krum, coordinate-wise median, trimmed mean) and pre-aggregation schemes that guarantee convergence under a bounded fraction of adversaries.
AI Safety & Alignment
Building on my internship, I aim to contribute to mathematically grounded safety: formal verification, robustness certificates, and interpretability tools that give us guarantees about model behavior rather than empirical hope.
Secure & Reproducible ML Engineering
Research tooling matters as much as theory. I build open-source, reproducible frameworks that are documented, tested, and packaged, so that robustness results can be re-run, attacked, and extended by anyone.
Learning Theory
From generalization bounds for two-layer ReLU networks to stochastic optimization under constraints, I enjoy the statistical and mathematical foundations that make robustness arguments rigorous.
Ongoing Work
M2 Research Internship at CMAP, Ecole Polytechnique
From April to October 2026, I work under El Mahdi El Mhamdi on robust distributed learning with adversaries: gradient manipulation attacks, open-source research tooling, and improvements to existing aggregation frameworks.
The experimental backbone is Krum, an open-source framework for Byzantine-resilient aggregation, installable via pip install krum.
PhD Directions
My PhD at CMAP, Ecole Polytechnique starts in November 2026, continuing this line of work. I plan to explore:
- Byzantine robustness at scale: aggregation rules that stay provably safe as clusters grow and decentralization increases.
- Formal verification for ML pipelines: bridging worst-case guarantees and practical training.
- Safety-critical applications: from federated learning to autonomous decision systems.
Talks
I present my work whenever I get the chance, and slides are linked when available.
Byzantine Robustness in Distributed Learning
September 2026: Internal Seminar, CMAP, Ecole Polytechnique. Introducing the Krum framework and our adaptive threshold mechanisms for Byzantine-resilient aggregation.
More talks and publications will appear here as my research progresses.