Get a reply from this employer
Your resume and cover letter, tailored to exactly what they’re hiring for.
Dex is seeking a highly experienced ML systems engineer to architect and operate real-time multimodal model serving at scale in London. You will own GPU utilisation, latency, and throughput optimisations across PyTorch, CUDA, Kubernetes on AWS/GCP, delivering robust production-ready platforms.
You will build scalable offline and online inference pipelines, transform research prototypes into maintainable platform features, and drive cost-efficient deployments.
This role is with one of Dex's trusted partner companies.
The role This company builds the multimodal data and evaluation platform powering model development for hundreds of leading AI teams, from autonomous vehicles to frontier robotics. Backed by top-tier global venture firms with significant growth funding, they provide the essential infrastructure that bridges raw visual and sensory data with production AI performance. You will join the in-house ML team sitting firmly on the engineering side, owning the architecture, deployment, and reliability of vision and multimodal models in production. This is not a role for prompt wrapping or tweaking hyperparameters in notebooks. It is a systems-first software engineering seat where you will solve hard GPU utilisation, latency, and throughput bottlenecks to deliver real-time model inference and closed‑loop data feedback at massive scale.
The work Architect and run low‑latency serving infrastructure for heavy vision and multimodal models, delivering real-time interactive predictions directly into high‑throughput annotation workflows. Build scalable offline inference pipelines processing millions of hours of egocentric video for robotics customers, balancing GPU cluster costs, throughput, and fault tolerance. Productionise agentic quality‑control and active learning workflows, transforming fragile research prototypes into resilient, self‑healing platform features. Design dynamic model integration layers that gracefully handle rapid, customer‑defined ontology shifts without requiring bespoke one‑off code. Profile and optimise end-to-end model serving performance across PyTorch, CUDA, and Kubernetes environments on AWS and GCP.