At LIX - Inria, I research privacy-preserving methods for LLM prompts. My work focuses on protecting prompts in LLM-based systems.
At Telecom SudParis (Institut Polytechnique de Paris), I led the design, implementation, and evaluation of ORION, a hierarchical multi-agent reinforcement learning framework where an LLM guides network-function placement across 6G domains.
At CNAM, I modeled LLM communication patterns and scheduling strategies for inference pipelines, and simulated distributed GPT-175B training on 1,024 GPUs. My research also includes federated fine-tuning of edge LLMs and low-resource NLP, where data augmentation improved BERT accuracy by over 10%.
