Post-Training Research Engineer

ABOUT BASETEN Baseten powers mission-critical inference for the world's most dynamic AI companies, like Cursor, Notion, OpenEvidence, Abridge, Clay, Gamma and Writer. By uniting applied AI research, flexible infrastructure, and seamless developer tooling, we enable companies operating at the frontier of AI to bring cutting-edge models into production. We're growing quickly and recently raised our $300M Series E , backed by investors including BOND, IVP, Spark Capital, Greylock, and Conviction. Join us and help build the platform engineers turn to to ship AI products. We are looking for an engineer with strong experience in machine learning and solid foundations in maths and computer science to join our growing Post-Training team at Baseten. Custom models are instrumental to the success of Baseten customers. By inference volume, the overwhelming majority of traffic at Baseten is to and from models that have been post-trained in some way, whether that be through reinforcement learning, supervised finetuning, a recent technique from the literature, or an in-house research technique from Baseten. The Post-Training team is responsible for the success of our customers’ post-trained models, and we employ a wide array of techniques to produce models that are more efficient and higher quality than even the biggest closed source models for the customer’s specific needs. Your role as a research engineer is to build the in-house tooling to support all of this. We care about training a wide spectrum of different model architectures with a variety of techniques efficiently and at scale. At times this involves zooming deep into a particular technical topic, but more often if involves working across the stack as a whole - systems-level concepts like Kubernetes, cgroups, storage systems, and networking topologies, as well as PyTorch distributed tensor computation, and GPU kernels. RECENT RESEARCH Dense, on-policy or both? Repeated kv cache for long-running agents Distillation without the dark – replicating black-box on-policy distillation on Baseten We don’t have a rigid set of skills, but here’s some of what we’re looking for: A deep understanding of modern ML techniques and tools for training transformers Advanced experience in a tensor/array computation library like PyTorch, TensorFlow, Jax, or similar A detailed understanding of transformer training parallelism strategies like data parallelism, sharded data parallelism, tensor parallelism, pipeline parallelism, context parallelism The experience and knowledge to profile and improve the performance of a distributed GPU program in PyTorch or a similar library The ability to perform roofline analysis on a transformer training setup A willingness to dive into messy problems, work with researchers, derive specifications by asking important questions, and execute Familiarity with HPC and distributed computing platforms like Slurm, Ray, Kubernetes, and Dask Familiarity with cluster networking technology like Infiniband, RoCE, GPUDirect Solid fundamentals in operating systems concepts like processes, files, kernel drivers, containerisation, and networking protocols A sense of creativity and willingness to ask difficult questions about our approach, assumptions, and tooling choices