Build the infrastructure
serious teams run on.

About the role

We're seeing more and more teams come to Aureus specifically for AI workload infrastructure — model training, fine-tuning runs, inference serving at low latency, vector search at scale. Right now we handle these like any other workload: great, but generic. We want to do better.

This role is for an engineer who has spent real time at the intersection of ML systems and production infrastructure — someone who understands not just how to run a GPU cluster, but why a training run falls apart at 90% completion, how to tune a vLLM deployment for throughput vs latency, or why your vector index rebuild is causing the rest of the cluster to degrade.

You'll work directly with customer teams to design their AI infrastructure, own the technical decisions on how we provision and manage GPU nodes, and help build the internal tooling that makes Aureus able to offer AI infrastructure at the quality level we hold everything else to.

What you'll do

• Design and operate GPU-backed clusters for training, fine-tuning, and inference workloads
• Advise customer engineering teams on AI infrastructure architecture — model serving strategies, vector DB selection, embedding pipeline design
• Build and maintain internal tooling for GPU cluster provisioning, utilisation monitoring, and cost attribution
• Define Aureus's standard configurations for common AI stacks (vLLM, Triton, Ollama, LangChain, LlamaIndex, Qdrant, Weaviate, pgvector)
• Collaborate with the platform team on Kubernetes scheduler configuration for GPU workloads (NVIDIA device plugin, MIG, time-slicing)
• Contribute to customer-facing documentation, runbooks, and security hardening guides for AI infrastructure
• Participate in on-call rotation for AI/ML customer clusters (shared with broader engineering team)

Requirements

• 5+ years of infrastructure or platform engineering experience
• 2+ years working directly with ML/AI workloads in production (not just experimentation)
• Hands-on experience with NVIDIA GPU clusters — provisioning, CUDA tuning, monitoring
• Strong Kubernetes background including custom schedulers, resource limits, node affinity
• Comfortable reading and reasoning about ML training code and inference pipelines even if you're not writing them
• Experience with at least one major vector database (Qdrant, Weaviate, Pinecone, pgvector)
• Strong written communication — we're remote, async-first, and documentation matters

Nice to have

• Experience with distributed training (PyTorch DDP, DeepSpeed, FSDP)
• Familiarity with model quantisation (GGUF, AWQ, GPTQ) and its infrastructure implications
• Background in MLOps tooling (MLflow, Weights & Biases, Kubeflow)
• Prior experience at an infrastructure company or managed services provider