SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Nebius is a Nasdaq-listed AI cloud infrastructure company headquartered in Amsterdam with R&D hubs across Europe, North America, and Israel. The company builds a full-stack AI cloud platform supporting developers and enterprises from data and model training through production deployment, with deep expertise in GPU orchestration, inference optimization, and applied AI systems.
The Nebius Token Factory is seeking a Senior Applied Scientist to lead research and optimization projects in efficient LLM and VLM inference. You will own focused research initiatives from hypothesis through experimentation, prototyping, and production handoff, translating frontier inference bottlenecks into rigorous research questions and actionable engineering solutions.
Your research will span quantization, quantization-aware training (QAT), distillation, speculative decoding, KV-cache optimization and compression, long-context inference, mixture-of-experts (MoE) routing, and model/runtime co-optimization. You will investigate how these approaches interact, evaluate trade-offs across model quality, latency, throughput, memory footprint, and cost per token, and collaborate with engineering teams to bring promising results into production.
Key responsibilities include:
- Leading research projects in efficient LLM/VLM inference from hypothesis through ablations, prototypes, and production handoff
- Developing and evaluating methods spanning quantization, QAT, distillation, speculative decoding, KV-cache reuse/compression, long-context inference, MoE routing, and model/runtime co-optimization
- Building prototypes using PyTorch, Triton, CUDA-adjacent tooling, and inference-serving frameworks; collaborating with MLEs and platform engineers to productionize components
- Researching LLM request routing and scheduling strategies, including cache-aware load balancing and prefill–decode disaggregation (PDD); evaluating how queueing, KV-cache transfer, worker placement, and capacity allocation affect latency, throughput, and serving cost
- Investigating multi-node inference for dense and MoE models, including wide expert parallelism (WideEP); studying expert placement, load imbalance, and computation/communication trade-offs
- Developing rigorous evaluations covering quality, latency, throughput, numerical stability, memory footprint, tail latency, and cost per token; comparing serving architectures under consistent conditions
- Working with MLEs, GPU kernel engineers, backend infrastructure, product, and customer teams to select research priorities with measurable production impact
- Sharing results through internal reports, technical blogs, papers, and open-source artifacts; mentoring engineers and scientists on experimental design, scientific rigor, and model/system trade-offs
REQUIREMENTS:
Must-haves:
- PhD in computer science, machine learning, ML systems, computer systems, computer architecture, electrical engineering, applied mathematics, or closely related discipline
- Strong publication record or equivalent research artifacts in ML, ML systems, efficient inference, model compression, quantization, distillation, serving systems, or related areas
- Strong Python and PyTorch implementation skills with ability to turn ideas into experiments and working prototypes
- Deep knowledge of LLMs, VLMs, transformer inference, decoding algorithms, model compression, quantization, and production-serving trade-offs
- Strong experimental design skills covering ablations, baselines, metrics, statistical reasoning, and failure analysis
- Excellent written and verbal communication
Nice-to-haves:
- First-author publications at venues such as NeurIPS, ICML, ICLR, MLSys, ACL, EMNLP, ASPLOS, OSDI, SOSP, ISCA, or HPCA
- Experience deploying ML models or inference optimizations in production
- Experience with vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo, FlashAttention, FlashInfer, Triton, CUDA, or PyTorch internals
- Experience applying post-training, SFT, DPO, RLHF, RLAIF, preference optimization, or synthetic data generation to inference quality or efficiency
- Open-source research artifacts, widely used benchmarks, technical blogs, or invited talks demonstrating contributions to efficient AI systems
- Research or implementation experience in distributed inference system architecture, such as LLM request routing, PDD, multi-node serving, or MoE expert parallelism including WideEP
