We grow our own verified talent

Become fluent across the major clouds, AWS, GCP, Azure, and OpenStack, and build, ship, and operate production infrastructure with an AI agent in the loop.

• Engineer on the big three clouds: core compute, networking (VPC/subnets/load balancers), storage, and IAM on AWS, GCP, and Azure
• Run your own private cloud on OpenStack (Nova, Neutron, Cinder, Keystone) and know when private beats public
• Provision and operate everything as code with Terraform, zero manual console clicks
• Run services on Kubernetes with health checks, limits, probes, and GitOps delivery
• Instrument systems with metrics, logs, traces, alerts, and SLOs
• Deploy and cost-control LLM serving and RAG infrastructure
• Build governed ops agents with human-approval gates

Model, move, and serve data with an AI agent in the loop, and build the data substrate AI systems run on.

• Model and load messy real-world data so assertion queries pass
• Build scheduled ELT pipelines with dbt tests green on an independent run
• Design cost-budgeted warehouse / lakehouse storage
• Enforce data quality with contracts and gates
• Build embedding and retrieval pipelines with measured quality

The advanced track. Go deep on the hardware, the parallelism, and the latency, the skills that get you hired at frontier labs and GPU clouds (the Nebius, NVIDIA, Anthropic, OpenAI tier).

• Reason below the framework: OS scheduling, memory, and concurrency, and read/write the C++/Rust that runtimes are built in
• Read the GPU stack: the CUDA execution model, SMs and warps, HBM bandwidth, and Tensor Cores
• Profile workloads top-down with PyTorch Profiler, Nsight Systems/Compute, and DCGM
• Scale training with data, tensor, pipeline, context, and expert parallelism (FSDP/ZeRO)
• Tune NCCL collectives and interconnects (NVLink, InfiniBand) so GPUs compute instead of wait
• Serve LLMs at low latency: prefill vs decode, KV cache, PagedAttention, continuous batching, vLLM / TensorRT-LLM / SGLang
• Run and acceptance-test GPU clusters on Kubernetes (GPU Operator, device plugin, DCGM dashboards)
• Cut cost and latency with quantization, mixed precision, and fused/custom kernels

The platform-at-scale track. Provision, network, secure, and operate fleets of GPU clusters across clouds and data centers, and serve models on them across providers. Built straight from how frontier labs run compute (think Anthropic Cluster Infra + Cloud Inference).

• Automate the full cluster lifecycle, provision, update, drain, recover, decommission, with agent-driven IaC (Terraform/Atlantis) and workflows (Temporal, Argo Workflows)
• Design multi-cloud networking: VPC design and peering, Shared VPC / Transit Gateway, Cloud Interconnect / Direct Connect, BGP and route control
• Build high-bandwidth, fault-tolerant cluster fabric that auto-drains and recovers on failure
• Master cluster + host networking: CNI (Cilium), eBPF, NetworkPolicy, multi-NIC, and service mesh (Istio/Envoy/Linkerd, mTLS)
• Make clusters secure-by-default: pod security standards, admission control, RBAC and least-privilege IAM, node/container hardening, image provenance
• Serve models across CSPs (AWS, GCP, Azure): request routing, capacity planning, autoscaling, and cost-aware placement to the cheapest accelerator and region
• Ship model versions to millions with CI/CD validation pipelines, and run real operational excellence: incident response, postmortems, on-call health