Bertrand Cabot

I help AI teams master their systems: optimization, benchmarking, qualification. LLM training, inference pipelines, scientific computing, comparing stacks or hardware: the central question is always the same — does the system do what it should, and how to prove it with numbers?

My approach combines two rare expertise in France:

→ Low-level GPU optimization: PyTorch / Nsight profiling, identifying regimes (math-bound, memory-bound, overhead-bound, SFU-bound), FSDP / DeepSpeed / vLLM tuning, 3D parallelism, Triton kernels.

→ System benchmarking and qualification: designing and executing test campaigns on AI infrastructure (training, inference, storage, multi-node), defining metrics (TTFT, ITL, throughput, scaling efficiency), comparing stacks (PyTorch vs JAX, vLLM vs TensorRT-LLM, FSDP vs DeepSpeed), LLM evaluation (Lighteval, MLPerf-style).

My background: 5 years at CNRS/IDRIS on the Jean Zay supercomputer (H100/A100/V100, InfiniBand, Lustre, SLURM). Expert support on national projects: BLOOM pretraining, OpenLLM France, Scribe, INESIA leaderboard, Hugging Face collaboration. 20 years of engineering including 10+ in Deep Learning and HPC.

What we can work on together:

→ GPU performance audit (5-10 days): complete diagnosis, bottleneck identification, quantified optimization plan

→ LLM training / inference optimization (long-term mission): FSDP, DeepSpeed, 3D parallelism, vLLM, TensorRT-LLM

→ System benchmark / qualification campaign (long-term mission): test plan design, cluster execution, technical reports for NVIDIA / clients / decision-makers

→ LLM and trustworthy AI evaluation: evaluation methods, hallucination analysis, reproducible benchmarks, leaderboard contributions

→ Training / skills transfer: distributed Deep Learning, GPU profiling, HPC usage