CV

PDF

Research interest

My research focuses on designing efficient and scalable systems for large-scale LLM applications. Specifically, my current topics cover LLM pre-training (e.g., automatic parallelization, heterogeneity, fault-tolerance) and serving agentic applications.

Education

• PhD in Computing, Imperial College London, 2025 - 2029 (expected)
  • Fully-funded PhD scholarship
  • Supervisor: Prof. Peter Pietzuch
  • Research topics: Systems for Machine Learning
• MSc. in Advanced Computing, graduated with distinction, Imperial College London, 2022 - 2023
• Diplôme d’Ingénieur, CentraleSupélec (Université Paris-Saclay), 2019 - 2023

Research

• Tangram: Hiding GPU Heterogeneity for Efficient LLM Parallelization
  Yanda Tao, Pedro Silvestre, Marcel Wagenlander, Peter Pietzuch, arXiv:2606.16907, under review.
  Poster presented at EuroSys’26.
  • A distributed LLM training system that decouples parallelization from mapping model partitions to heterogeneous GPUs, to achieve efficient parallel training in a wide range of scenarios. Tangram achieves up to 2.1× training throughput compared to existing heterogeneous parallelizers by integrating state-of-the-art homogeneous parallelizers.
• MegaLog: Heterogeneous Fault Tolerance for Efficient LLM Training
  Pedro Silvestre, Yanda Tao, Marcel Wagenlander, Otto White, Peter Pietzuch, 2026, under review.
  • An investigation of existing checkpointing- and redundancy-based fault-tolerant approaches to LLM training. Our solution, MegaLog, minimizes both runtime overhead and recovery time through fine-grained training state management. MegaLog provides 34% higher effective training time even under frequent failures, and 4× faster recovery time.
• Scepsy: Serving Agentic Workflows Using Aggregate LLM Pipelines
  Marcel Wagenländer, Otto White, Britannio Jarrett, Pedro Silvestre, Yanda Tao, Guo Li, Huanzhou Zhu, Lluís Vilanova, Peter Pietzuch, arXiv:2604.15186, under review.
  • A new agentic serving system that efficiently schedules multi-LLM agentic workflows by exploiting fractional GPU shares and aggregate workload predictions as a pipeline. Scepsy achieves up to 2.4× higher throughput and 27× lower latency compared to existing agentic serving systems.

Experience

• Research Assistant, SysML, LSDS, Imperial College London, 2024
  • Samsung-funded position
  • Investigated automatic parallelization and fault tolerance in LLM training.
  • Our investigation led to the development of Tangram and MegaLog, which are currently under review.
• Junior Developer, SAP France, 2022
  • Built a patch-dependency verifier that gated new releases of SAP commercial products, including BusinessObjects BI tools.
• Data Scientist/MLOps Intern, Alstom Transport, Data Science Team, 2021 - 2022
  • Provided operational support to the radio log analysis project (featured in press release)
  • Worked on AI pipeline improvement: 10× speed-up using optimized data structures and transfers
• Member of technical team, ViaRézo, 2022 - 2023
  • Project Webaz: software infrastructure for verifying subscription and maintaining network access in a campus of 2000+ students.

Skills

• Programming: Python, Java, C++, Bash, JavaScript (Node.js)
• Tools: GitHub, Docker, Kubernetes, AWS, Azure, IaC tools (Ansible/Terraform)
• Frameworks: PyTorch, Megatron, DeepSpeed, NCCL, JAX/XLA, Ray
• Languages: French native speaker, English (IELTS 8.0, TOEFL iBT 109/120), Spanish (B1)

Hobbies

• Leader of CentraleSupélec’s choir (LyriCS), 2019 - 2021
• Vice-President of CentraleSupélec’s board game society, 2020 - 2021