Hybrid LLM Inference Benchmark Engineer

CommonAI C.I.C.; Cambridge; Hybrid; GBP 60,000 - 90,000; Full time

Benefits offered by this job

Pension

Professional development opportunities

Job summary

CommonAI CIC invites engineers to join the Scaling Inference Lab in Cambridge to build reproducible LLM serving workloads and benchmarks. You will connect application behavior to physical measurements and the lab’s simulation platform, focusing on long-context inference and agentic workloads.

You will work with measurement and simulator engineers to preserve workload behavior across real systems and models, and document configurations, provenance and usage for reproducibility.

Qualifications

• Experience building or evaluating ML systems, inference services or comparable workload-driven systems.
• Strong software skills for constructing reliable benchmark tools and debugging interactions.
• Understanding of LLM inference and how serving choices affect performance and output quality.
• Experience with AI coding agents and validating generated code or data transformations.

Responsibilities

• Build reproducible LLM serving workloads and benchmark tools, including long-context workloads.
• Characterise request patterns, lengths, caching, batching, and tool interactions.
• Define performance measures alongside application-quality requirements.
• Collaborate with measurement and simulator engineers to reproduce workloads on physical systems.
• Investigate how workload changes affect conclusions and keep benchmarks relevant.

Skills

ML systems experience

Benchmark tooling

LLM inference

Agentic workloads

Code validation

Tools

vLLM

SGLang

Job description

CommonAI CIC invites engineers to join the Scaling Inference Lab in Cambridge to build reproducible LLM serving workloads and benchmarks. You will connect application behavior to physical measurements and the lab’s simulation platform, focusing on long-context inference and agentic workloads.

You will work with measurement and simulator engineers to preserve workload behavior across real systems and models, and document configurations, provenance and usage for reproducibility.