Forward Deployed Engineer, Labelbox

• 2+ years of professional software engineering experience, with strong fundamentals in Python and at least one systems-level language (Go, Rust, C++)
• Demonstrated experience with containerization and sandboxing (Docker, Podman, Firecracker, or similar) in production or near-production contexts
• Familiarity with RL concepts: MDPs, reward shaping, episode structure, observation/action spaces. You don’t need to have trained models, but you need to understand what an environment must provide to an RL training loop
• Experience building or maintaining developer tooling, CLI tools, or infrastructure automation
• Comfort working with browser automation frameworks or terminal interaction tooling
• Strong debugging instincts—you can trace failures across process boundaries, container layers, and network calls
• Ability to read and implement from academic papers and open-source benchmark repositories without extensive hand-holding
• The ideal candidate is a strong software engineer first, with genuine curiosity and working knowledge of reinforcement learning
• You’ve probably built infrastructure or developer tooling at a startup or mid-stage company, and you’ve been pulled toward the ML/AI space—maybe through side projects, open-source contributions, or a prior role adjacent to an ML team
• You’re the kind of engineer who reads an RL benchmark paper and immediately thinks about how to make the environment more robust, not how to improve the policy gradient
• You thrive in ambiguity
• You can take a loosely defined project requirement—“build an environment that tests an agent’s ability to navigate a file system and execute multi-step bash workflows”—and deliver a working, tested, documented system without needing a detailed spec
• You move fast, but you care about reliability because you know environments that break silently poison training data

Desirable

• We’re hiring a Forward Deployed Engineer to own the design, development, and operationalization of reinforcement learning environments
• You’ll build the sandboxed, reproducible execution environments that AI agents interact with during training and evaluation—things like terminal-based task benchmarks, browser and computer-use environments, and tool-augmented agentic workspaces
• This is a hands-on engineering role
• You’ll write production-quality infrastructure code, integrate with open-source RL tooling, and work closely with our data operations team to ensure environments are robust, observable, and ready for human annotators and model agents alike
• You won’t be doing ML research, but you’ll need to deeply understand how RL training loops consume environments and where the bottlenecks live
• You'll design, build, and maintain sandboxed RL environments for agentic AI training—including terminal emulators, browser automation harnesses, computer-use simulators, and tool-augmented workspaces (e.g., environments built on frameworks like TerminalBench, OSWorld, and Tau-bench)
• Develop reproducible, containerized execution environments (Docker, VMs, lightweight sandboxes) that support deterministic task rollouts and reward signal collection
• Integrate with and extend open-source agentic tooling and custom CLI/API harnesses to enable multi-step agent interaction
• Build instrumentation and observability layers—structured logging, trajectory capture, state snapshotting—so training runs and human annotation sessions produce clean, auditable data
• Collaborate with data operations to design task curricula and evaluation protocols that stress-test model capabilities across environment types
• Own environment deployment and reliability: CI/CD pipelines, automated testing of environment configurations, and monitoring for drift or breakage across versions
• Rapidly prototype new environment types as client and internal requirements evolve, moving from spec to working system in days, not weeks
• You’ll work across a portfolio of projects spanning different AI labs and model capabilities—no single-product monotony. The environment types you build will evolve as the frontier of agent capabilities moves