RL Environment Engineer (Full-time, Part-time, or Contract)

About the role

Besimple AI is building datasets, benchmarks, and training environments that help voice models understand people, interact naturally, and complete real tasks.

We’re looking for an engineer with experience building reinforcement learning environments to help us create the next generation of conversational-agent training systems. You’ll turn real-world interactions into tasks that agents can practice, receive feedback on, and learn to complete reliably.

Experience building RL environments is the core requirement. Speech experience is a plus. We welcome backgrounds in language agents, coding, computer use, robotics, games, simulation, and other RL domains.

What you’ll work on

• Build training environments. Define tasks, observations, actions, state transitions, rewards, and termination conditions. Create environments where agents ask questions, use tools, recover from errors, and reach verifiable outcomes.
• Design rewards and verifiers. Implement checks for task completion and intermediate progress. Find reward-hacking opportunities, ambiguous success criteria, and shortcuts that let agents pass without solving the task.
• Turn human interactions into tasks. Work with recorded conversations and human demonstrations to construct scenarios, starting states, alternative continuations, and realistic failure cases.
• Build interactive simulators. Develop user and tool simulators that respond to an agent’s actions, maintain consistent state, and support varied, repeatable interactions.
• Make environments reliable at scale. Build rollout interfaces, parallel execution, logging, replay, and debugging tools. Make failures reproducible and environments easy to connect to training pipelines.
• Validate what agents learn. Work with researchers to run baseline policies and training experiments, measure learning progress, and test whether improvements transfer to unfamiliar scenarios.

Our initial focus is voice agents: systems that need to listen, speak, call tools, handle interruptions, and finish tasks while interacting with people. You’ll help determine how these behaviors should be represented and measured in an RL environment.

What we’re looking for

• Hands-on experience building or substantially extending environments used to train or evaluate RL agents.
• Strong engineering skills and experience writing tested, maintainable software.
• Experience defining task interfaces, managing environment state, implementing rewards, and debugging agent–environment interactions.
• The ability to investigate why an agent fails—or succeeds for the wrong reason—and improve the environment accordingly.
• Clear technical communication and comfort taking an open-ended problem from an initial prototype to a reliable system.

We welcome recent graduates with substantial academic or open-source projects, as well as engineers with industry experience. A research publication is not required; working environments, code, and thoughtful experimental results are useful evidence.

Especially relevant experience

• Speech, audio, conversational agents, or full-duplex interaction.
• Tool-use, coding, browser, or computer-use environments.
• User simulators, executable rewards, reward models, or preference data.
• Gymnasium-style interfaces, distributed rollouts, or integration with RL training frameworks.
• Using human demonstrations or real-world interaction data to construct training tasks.
• Reproducible evaluation, deterministic replay, and testing for reward hacking.

Why this role

You’ll own a core part of how our models learn: the tasks they encounter, the feedback they receive, and the conditions under which we measure progress.

This is an opportunity to shape an early engineering and research direction, build environments grounded in real human interactions, and see how your work changes the behavior of conversational agents.

About Besimple AI

Why Us

At Besimple AI, we’re making it radically easier for teams to build and ship reliable AI by fixing the hardest part of the stack: data. Good evaluation, training and safety data require domain experts, robust tooling and meticulous QA. AI teams and labs come to us to get high quality data so they can launch AI safely. We’re a YC X25 company based in Redwood City, CA, already powering evaluation and training pipelines for leading AI companies across customer support, search, and education. Join now to be close to real customer impact, not just demos.

Why This Matters

High-quality, human-reviewed data is still the single biggest driver of model quality, but most teams are stuck with old tools and legacy processes that do not scale to modern, multimodal, agentic workflows. Besimple replaces that mess with instant custom UIs, tailored rubrics, and an end-to-end human-in-the-loop workflow that supports text, chat, audio, video, LLM traces, and more. We meet teams where they are—whether they need on-prem deployments and granular user management or a fast cloud setup—to turn evaluation into a continuous capability rather than a one-time project.

Traction & Customers

Who You’ll Work With

Founders previously built the annotation platform that supported Meta’s Llama models. We’ve seen how world-class annotation systems shape model quality and iteration speed; we’re bringing those lessons to every AI team that needs to ship with confidence. You’ll work directly with the founders and users, owning problems end-to-end—from an interface that unlocks a tough rubric, to a workflow that reduces disagreement, to a AI judge system that improves quality.

How We Work

• Bias to shipping and learning with customers
• Respect for craft: calibration, rubric clarity, inter-annotator agreement (IRR)
• Tight feedback loops from production back to evaluation
• Ownership: you’ll shape evaluation as an engineering discipline with real “fail-to-ship” tests tied to business and safety goals

If you’re excited by systems that combine product design, human judgment, and applied AI—and you want to build the data and evaluation layer that keeps AI trustworthy—come build with us. See how fast teams can go from raw logs to a robust, human-in-the-loop eval pipeline—and how that changes the way they ship AI.