Member of Technical Staff | ML Systems

Role overview Build the internal machine learning systems that take research candidates from data preparation and training through evaluation to governed, reproducible releases. The platform serves researchers and engineering teams, and must make it possible to run and audit models across cloud and customer environments. The work combines distributed training, data infrastructure, experiment tooling, and release governance.

Responsibilities - Develop compute primitives and improve training and serving performance for graph neural networks. - Evolve distributed sampling and training capabilities, including neighbor sampling. - Define binary data formats and support data materialization and feature backfills for training and evaluation. - Establish data contracts with teams producing customer and proprietary datasets. - Build and operate experiment tracking, checkpointing, and evaluation systems that support reproducibility. - Maintain model registry, lineage, versioning, compatibility, and release gates; ensure releases can be traced to their data, code, configuration, and evidence.

Requirements - Strong systems engineering skills and experience with multi-node GPU workloads and distributed training frameworks such as PyTorch Distributed. - Experience with columnar data formats and large-scale data materialization. - Familiarity with experiment tracking, model registries, evaluation, and reproducibility across the ML lifecycle. - A product mindset for internal platforms and their users. - Data science experience is not required.

Nice to have - CUDA kernel development or GPU performance optimization. - Graph neural networks or graph sampling at scale. - Lance, Arrow, or other columnar or indexed storage formats. - Multi-cloud GPU compute, model governance, or audit requirements in regulated environments.

Benefits and work setup Full-time, remote role listed for São Paulo. Success is measured in part by faster experimentation and governed releases, higher training throughput per GPU, and complete lineage and reproducibility for production models.