Staff Software Engineer, RL Environments

About Scale AI At Scale, our mission is to develop reliable AI systems for the world's most important decisions. Our products provide the high-quality data and full-stack technologies that power the world's leading models, and help enterprises and governments build, deploy, and oversee AI applications that deliver real impact. Scale Frontier Data is the organization behind the training and evaluation data that frontier labs depend on. We build the systems, tooling, and expert workflows that turn hard human expertise into signals that models can learn from, across reasoning, coding, agentic tool use, and domain expertise. Reinforcement learning environments are now the center of gravity for that work: the difference between a model that demos well and a model that reliably completes long-horizon work is almost always the quality of the environments and reward signals it was trained against. Responsibilities As a Staff Software Engineer, RL Environments, you'll own the technical foundation for how Scale builds, runs, verifies, and delivers RL environments at scale. An RL environment is a real piece of software: a containerized world with real dependencies, real state, real tools, and a grader that has to be correct even when the agent is creative about breaking it. Building one is a full-stack engineering problem. Building thousands of them reproducibly, cheaply, with trustworthy reward signals and throughput measured in millions of rollouts is a systems problem that very few people have solved. You'll work on both. You'll design the platform: sandboxed execution, environment packaging and versioning, rollout orchestration, trajectory capture, verifier frameworks, and the authoring surfaces that let engineers and domain experts produce environments without reinventing infrastructure each time. And you'll go deep on the environments themselves by instrumenting real applications, designing task suites that expose specific capability gaps, and building graders that hold up under adversarial optimization. This is a hands-on engineering role. You'll set technical direction across multiple teams, and you'll still be the person who writes the hard part. Required Qualifications 8+ years of software engineering experience with strong fundamentals in distributed systems, system design, data structures, and algorithms. Strong Python skills and a track record of shipping production software; comfort in at least one other part of the stack (TypeScript/React, Go, Rust, or similar). Deep experience with containerization and sandboxed execution, including Docker, VMs, gVisor/Firecracker, Kubernetes, or equivalent. Experience building or operating high-throughput backend systems: orchestration, job scheduling, queuing, and large-scale data pipelines. Hands-on experience building with LLMs including agent loops, tool calling, MCP, or eval harnesses, and enough intuition about model behavior to reason about what a training signal actually teaches. Demonstrated ability to own ambiguous, undefined problems end to end and drive them to a shipped system. Excellent written and verbal communication; ability to align engineers, researchers, and non-engineering partners on a technical direction. Preferred Qualifications RL & Post-Training Direct experience building RL environments, agentic benchmarks, or eval harnesses (SWE-bench-style task suites, terminal or browser environments, tool-use benchmarks, or in-house equivalents). Familiarity with post-training methods: RLHF, RLAIF, RLVR, GRPO/PPO-family algorithms, rejection sampling, reward modeling, and the practical failure modes of each. Experience designing verifiable reward signals, and firsthand experience with reward hacking and how to defend against it. Experience with RL training or serving stacks (verl, TRL, Ray, vLLM, SGLang, or similar). Systems & Infrastructure Experience with high-scale sandbox or code-execution infrastructure, remote development environments, or CI systems. Experience with cloud-native infrastructure across AWS/GCP/Azure, Infrastructure as Code, and CI/CD. Strong observability instincts: tracing, structured logging, and metrics for systems whose failure modes are statistical rather than binary. Experience building internal tools that non-engineers rely on daily, especially data-dense review and annotation interfaces. Ways of Working Experience in a research-adjacent engineering role, translating research goals into production systems. Experience working directly wi