About AZX

Our mission is to accelerate positive impact in critical industries through AI transformation. We specialize in physics-informed ML and enterprise AI solutions that directly address climate and sustainability challenges.

We’re growing quickly and already work with category-leaders in real estate (CBRE), energy (LevelTen Energy), logistics (Flexe) and utilities.

We bootstrapped profitably for our first year and are now backed by leading investors focused on AI, climate and energy.

We work on challenges in clean energy, decarbonization, climate risk, energy systems, and global economics. We’re building our company for long-term success and aim to create the ultimate place to work for those passionate about AI and making a positive impact.

About This Role

You will be responsible for owning the layer where AI work physically happens: the machines, the isolation boundary, and the models running on them. This role anchors on two systems. The first is our inference control plane — open-weight models and custom task-model zoos, hosted and operated across managed GPU clouds and customer-managed Kubernetes clusters, with scale-to-zero economics, cold-start discipline, and per-token cost accounting that stays correct even when a client disconnects mid-stream — along with the Kubernetes layer those workloads live on: operators, autoscaling, node lifecycle. The second is our agent-sandboxing platform: hardware-isolated microVMs for running untrusted, agent-generated code securely and compliantly by construction, where agents operate with least privilege, never see a credential, and a human gates anything that writes to a system of record. You'll write Rust in the morning, a Kubernetes controller after lunch, and a FastAPI control-plane endpoint before you go home — building the fork engine, the guest agent, and the multi-substrate model lifecycle. We're looking for individuals who've built this class of stack (an inference-serving or serverless-GPU platform), operated it hard at scale, or ideally both.

Responsibilities

Manage the serving tier for open-weight models: engine deployment and configuration, cold-start strategy, per-model SLOs, and upgrade/canary discipline.

Administer the Kubernetes layer for inference and sandbox workloads: operators and CRDs, autoscaling (KEDA/Karpenter-class), GPU scheduling and sharing, and node lifecycle.

Own the stateful data plane end to end with restore procedures that are regularly tested.

Oversee the sandbox runtime and its host-side control plane: lifecycle, exec, snapshot/fork, teardown, metering, and the threat model of the isolation boundary.

Direct the FastAPI control-plane services, Terraform/OpenTofu, Bicep, and the dashboards

Run the layer the backend services team builds on, expect to debug into their services, and expect them to read your dashboards.

Manage the open-source posture: build to OSS standards and release as it matures, with reviewed PRs, real docs, and reproducible builds.

Core Qualifications

5+ years of shipping production systems in a systems language. Rust is the house language, but polyglots are welcome — deep Go, C/C++, or Zig with genuine appetite for Rust counts. Async runtimes, memory-safety discipline, and debugging at the syscall boundary should be familiar territory.

Operated Kubernetes workloads that other people depended on — controllers or operators, scheduling, autoscaling, node lifecycle. You've been paged, and the experience changed how you build.

Strong ability to threat-model isolation boundaries (namespaces, cgroups, seccomp, hypervisors), including identifying what an untrusted guest could observe, forge, or exhaust, and applying security best practices for agentic execution — least privilege, no credentials in the sandbox, audit trails, and human approval on write actions.

Hands-on experience deploying or operating open-weight LLM serving infrastructure (vLLM/SGLang or similar), including packaging models into reliable, metered production endpoints.

Performance discipline in distributed systems: you measure before you optimize, and you can tell the story of a latency you killed with the numbers attached.