Machine Learning Engineer — Distillation

Role overview Develop methods and production systems that make large models smaller, faster, and less expensive to serve while preserving useful output quality. You’ll own work across research and engineering, from designing distillation pipelines and running experiments to delivering models and supporting tools for production use.

Responsibilities - Design and implement knowledge-distillation approaches, including teacher–student, self-distillation, and multi-teacher methods. - Distill foundation models into smaller models suited to lower-cost, faster inference. - Run and analyze large training experiments, comparing quality, latency, and cost. - Turn new research ideas into production-ready implementations in collaboration with research colleagues. - Improve training and inference performance, including memory use, throughput, and latency. - Build or improve internal tooling, evaluation frameworks, and experiment tracking; open-source contributions may also be part of the work.

Requirements - Strong background in machine learning or deep learning. - Hands-on experience with distillation for language models or other neural networks. - Solid understanding of training dynamics, loss functions, and optimization. - Experience with PyTorch or JAX and current machine-learning tools. - Comfort running experiments on multi-GPU or distributed systems. - Ability to assess tradeoffs between model quality and performance, with a focus on shipping usable systems.

Nice to have - Experience distilling LLMs or other large sequence models. - Inference optimization experience, such as quantization, pruning, or kernel work. - Familiarity with language-model evaluation, research publications, or open-source contributions. - Experience in an early-stage or fast-moving team.

Benefits and work setup - Remote-friendly, asynchronous work environment. - High ownership and direct influence on product and technical direction. - Small team combining research and engineering; competitive compensation and equity are noted, with no figures provided.