Nicolas Castet is a Senior Deep Learning Engineer with 10 years of experience applying optimization, parallel computing, and ML research to production-grade systems, currently building at NVIDIA in Austin. He brings full product lifecycle expertise—from algorithm design and correctness/performance testing to DevOps for distributed training—backed by graduate research and engineering degrees from EPF and Texas A&M. His open-source work includes substantive backend and CI/CD contributions to Horovod and distributed-compute improvements in JAX, adding cluster integrations (SLURM, Open MPI, LSF) and GPU allreduce support that improved reliability on multiple platforms. Comfortable in Modern C++ and Python, he focuses on making large-scale deep learning systems both faster and more robust. An understated strength is his history of refactoring build systems and initializing complex multiprocess GPU tests, bridging research code with production deployment.
10 years of coding experience
6 years of employment as a software developer
Master, General Engineering; Computer Science, Master, General Engineering; Computer Science at EPF Ecole d'Ingénieur-e-s
Master of Science, Computer Science, 3.5, Master of Science, Computer Science, 3.5 at Texas A&M University
Distributed training framework for TensorFlow, Keras, PyTorch, and Apache MXNet.
Role in this project:
Back-end & DevOps Engineer
Contributions:64 reviews, 27 commits, 34 PRs in 3 years 9 months
Contributions summary:Nicolas primarily contributed to the backend of the Horovod framework, adding support for IBM PowerAI DDL GPU allreduce and Spectrum MPI. They refactored the build process by integrating native libraries into the CMake build system. Additionally, they addressed several issues related to build failures and deprecated the DDL backend while improving the core functionality of the framework. The user also improved the performance and reliability through modifications in the CI/CD, and added support for LSF job management.
Composable transformations of Python+NumPy programs: differentiate, vectorize, JIT to GPU/TPU, and more
Role in this project:
DevOps Engineer & ML Engineer
Contributions:37 reviews, 4 commits, 12 PRs in 5 months
Contributions summary:Nicolas primarily focused on improving the JAX distributed system, fixing bugs and adding features related to distributed computing. They fixed an issue in `MultiProcessGpuTest` involving incorrect execution and initialization within the distributed environment. They also increased the timeout for distributed initialization, added a generic interface for automatic initialization, and added support for Slurm and Open MPI cluster environments. Furthermore, the user fixed a bug in the Mosaic GPU implementation.
pytorchpythonjitautomatic-differentiationgpu
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.
Request Free Trial
Nicolas Castet - Senior Deep Learning Engineer at NVIDIA