Senior Applied Scientist: Efficient LLM Inference & Optimization

FDE Times; Dubai; On-site; AED 717,000 - 963,000; Full time

Benefits offered by this job

Health insurance

401(k)

Parental leave

Remote work stipend

Job summary

Nebius is seeking a scientist to turn frontier LLM and VLM inference bottlenecks into research questions, conduct rigorous experiments, and translate findings into production capabilities. You will own focused projects, write code, and publish technical artifacts for engineers to build on.

Your work spans quantization, distillation, speculative decoding, KV-cache optimization, and model/runtime co-optimization, collaborating with engineering teams to deploy results in production.

Qualifications

• PhD in CS/ML or related field is required.
• Strong Python and PyTorch implementation skills.
• Deep knowledge of LLMs/VLMs, transformer inference, quantization.
• Excellent experimental design and communication skills.

Responsibilities

• Lead research projects in efficient LLM/VLM inference from hypotheses to production handoff.
• Develop and evaluate methods: quantization, QAT, distillation, KV-cache optimization, and co-optimization.
• Build prototypes using PyTorch, Triton, CUDA-adjacent tooling, or inference-serving frameworks.
• Research LLM routing, scheduling, and cache-aware load balancing; evaluate latency, throughput, and cost.
• Investigate multi-node inference, wide EP, expert placement, and performance trade-offs.
• Provide rigorous evaluations on quality, latency, throughput, memory, stability, and cost per token.
• Collaborate with MLEs, GPU kernel, backend, product, and customer teams; set priority with measurable impact.
• Share results via reports, blogs, papers, and open-source artifacts; mentor teammates.

Skills

Python

PyTorch

LLMs

VLMs

Research methods

Education

PhD in CS/ML

Tools

Triton

CUDA

Torch internals

Job description

Nebius is seeking a scientist to turn frontier LLM and VLM inference bottlenecks into research questions, conduct rigorous experiments, and translate findings into production capabilities. You will own focused projects, write code, and publish technical artifacts for engineers to build on.

Your work spans quantization, distillation, speculative decoding, KV-cache optimization, and model/runtime co-optimization, collaborating with engineering teams to deploy results in production.