About Corvex

Corvex is a neocloud building GPU infrastructure for AI training and inference. We run 100% Kubernetes-native, delivering dedicated or shared GPU clusters with isolated networking, managed storage, and automated provisioning. Customers run production training and inference on our hardware and hold us to hyperscaler standards.

We are building the platform to stamp out new datacenters from automation: rack the servers, run one command, platform is live.

We are a small team with a large mandate and we hire accordingly.

The Role

We need an experienced, top-level architect to own two things at once: the cross-domain architecture that turns nine parallel POCs into one platform, and a primary domain where you are the architect of record and write production code. We don't want a pure coordinator. The platform comes together because you are in the codebase, not above it.

Nine engineering domains are running in parallel: platform bootstrap, networking, bare metal provisioning, Kubernetes, storage, observability, controls and billing, performance testing, and security. Eight are domain-scoped; security is cross-cutting and touches every other domain. Each has a dedicated lead who owns their domain’s architecture. Your job is to make sure the full system holds together, that decisions made in networking don’t break Kubernetes, that the provisioning pipeline hands off cleanly to the K8s platform, and that nine domain POCs converge into one coherent, delightful, and industry leading product.

You will also push the team forward with AI tooling and agents, using LLMs, code generation, and agentic workflows to build faster. We are a small team tackling a large-scope platform. The right architect does not just design systems; they accelerate the team by finding leverage through modern tooling.

The platform is greenfield today. We need to run POC to prove the architecture end-to-end before scaling into production across multiple sites. You will arrive during the period that defines how this platform works for the next decade.

What you’ll own

Architecture and integration

• Own the overall vision and E2E technical integration across the nine domains. You are the connective tissue between domain leads who each own their own architecture.

• Drive architectural coherence ensuring cross-domain interfaces (bare metal provisioning to K8s join, VPC to tenant isolation, observability to control plane) are designed once, not discovered late.

• Identify architectural risks early, DPU MVP migration path, cross-tenant blast radius, BMC credential lifecycle, GPU failure handling at scale, multi-site control plane topology.

• Architect to externally published SLAs, boot time, API uptime, RDMA point-to-point ib_bw_write, GPU failure rates.

• Participate in key design sessions, VPC-to-K8s integration, tenant API orchestration model, networking abstraction shim architecture, observability pipeline design, storage plane selection, two-tier dedicated/shared K8s tenancy.

• Review all domain plans and flag gaps, conflicts, or missing integration points before they become production problems.

Vendor evaluation

• Support vendor evaluations across networking, bare metal provisioning, Kubernetes, and controls.

• Help domain leads assess how vendor choices compose into a full platform. We evaluate combinations, not vendors in isolation.

• Bring the technical depth to ask the questions vendors don’t want to answer.

• Drive the build-vs-buy call on each layer with evidence from hands-on evaluation, not vendor decks.

AI tooling and acceleration

• Drive adoption of AI-assisted development across the team, code generation, agentic coding workflows, AI-powered debugging and review.

• Identify where LLMs and agents can replace manual engineering effort, IaC generation, test creation, documentation, operational runbooks.

• Build or integrate agentic workflows into CI/CD, provisioning, and platform operations.

• Set the standard. The team should see you using these tools effectively and adopt from your example.

Scope boundaries

• You own a primary domain and the cross-domain architecture. You are the architect of record for one of the nine domains, the one where your depth runs strongest. The other eight have dedicated leads who own their decisions; you partner with them, challenge them, and connect their work into one platform making good architectural decisions.

• You won’t program manage. A program manager owns tracking and delivery. You focus on technical architecture.

• You own cross-domain decisions. Domain leads decide within their stack. You decide the interfaces, shared patterns, and how the pieces compose into a platform, and you have the obligation to escalate when coherence is at risk.

Working with the Team

• You join an established engineering team.

• You partner with domain leads as peers, not reports. Influence comes from technical depth, sound judgment, and the trust you build. You manage the technology.

What we’re looking for

Required

• 12+ years in infrastructure engineering, with at least one tour as the architect for a multi-tenant platform serving external customers.

• Designed or operated a multi-tenant infrastructure platform at scale, cloud provider, neocloud, GPU cloud, HPC, or large-scale Kubernetes platform.

• Hands-on depth across the full stack: bare metal provisioning, networking (EVPN/VXLAN, InfiniBand or RoCE), Kubernetes internals, and storage (Ceph, VAST, Weka, or similar).

• Can go deep in at least three of: networking, Kubernetes, bare metal provisioning, storage, GPU/HPC.

• Has made vendor build-vs-buy decisions with real data from hands-on evaluation, not from vendor marketing or analyst reports.

• Has integrated systems across team boundaries where no one person owned the full picture.

• Has operated in a greenfield environment where the architecture wasn’t handed to you.

• Actively uses AI coding tools (Claude, Copilot, Cursor, or similar) as a core part of how you work, not as a novelty. Can articulate where they help and where they don’t.

Strongly preferred

• Bare metal depth: IPMI/Redfish, PXE, Metal3/Ironic, firmware lifecycle management.

• Deep Kubernetes: Cluster API, building custom operators, multi-cluster management (Kamaji, vCluster, or similar), multi-tenancy patterns.

• Networking at the fabric level: EVPN/VXLAN, Netris or similar SDN, multi-vendor switching (NVIDIA Spectrum, Drivenets, Arista, Broadcom).

• GPU infrastructure: NVIDIA GPU Operator, DCGM, NCCL/RCCL, MIG, InfiniBand fabric management, AMD ROCm stack.

• Has shipped a platform to external paying customers, not just internal tooling.

• Has integrated AI/ML workload requirements into platform design, understands what training and inference jobs actually need from infrastructure.

• Has built or integrated agentic workflows into real engineering systems (CI/CD, provisioning, IaC, runbooks).

• Familiarity with industry benchmarks for GPU clouds (e.g., SemiAnalysis ClusterMAX) and what it takes to clear them.

Not a fit if

• Pure hypervisor/VMware background with no bare metal experience.

• Only operated existing platforms and never designed one from scratch.

• Needs a fully defined architecture before contributing.

• Views AI tooling as hype rather than a practical multiplier.

• Don’t want to own and deliver.

How we work

• We are small, experienced, and move fast. Domain leads have real authority over their domains; your job is to make sure what they build works as a system.

• We do not gate decisions through committees. We have explicit owners, written designs, and rigorous reviews. Disagreement is expected; unresolved disagreement is not.

• Customers run production on our platform today. Architecture quality shows up in incidents, cost, and time-to-next-site.

AI as a force multiplier

Corvex has made AI-assisted engineering a first-class expectation across infrastructure, not an opt-in. We use Claude Code and MCP servers against our internal systems (Jira, NetBox, CheckMK), and we are converting tribal infrastructure knowledge into AI-consumable formats (CLAUDE.md / AGENTS.md) so engineers and agents can operate against the same source of truth.

We expect this role to raise that bar. If you cannot point to concrete ways you have used these tools to ship faster, this is not the right seat.

What success looks like

Through the POC (Weeks 1–11)

• Week 4: Cross-domain interface contracts are documented for all nine domains. Integration risks are identified and owned.

• Week 8: Vendor evaluations have converged on a recommended stack with evidence and trade-offs clearly captured. Integration dry-runs have exposed the hard problems.

• Week 12: POC demonstrates end-to-end provisioning of a tenant cluster on automation, with observability, isolation, and billing hooks in place. A production architecture document exists and the team aligns on it.

First 6 months

• Production rollout of the platform architecture on the new site is underway, with clear hand-offs between domains.

• AI-assisted workflows are in daily use across at least three domains, IaC generation, runbook authoring, and automated review.

• A “site stamp-out” playbook exists: the architectural steps required to stand up the next datacenter are written down, reviewed, and tested against at least one dry-run.

First 12 months

• A second site can stand up against the platform in materially less time than the first, with the architecture holding up to external customer load.

• Architecture decisions are documented, durable, and hold up to outside review (auditors, enterprise customers, security assessments).

• The team ships faster with the same headcount than it did when you joined, and the AI tooling footprint is a measurable reason why.