Member of Technical Staff — Agentic Model Training

Introducing Moonlake, AI for creating world simulations.

About Moonlake

Moonlake is building the frontier of interactive world models: systems that generate, simulate, and reason over 3D environments for robotics, embodied AI, and interactive applications.

We develop the infrastructure that enables intelligent systems to learn, evaluate, and interact within realistic virtual environments before operating in the physical world.

Our work sits at the intersection of:

• Robotics

• Embodied AI

• Interactive 3D Worlds

• World Models

• Simulation Infrastructure

• Physical AI

Moonlake is building the next generation of AI infrastructure for interactive digital worlds. Our mission is to enable anyone to create, simulate, and interact with rich environments using natural language and multimodal inputs, turning simple ideas into worlds with structure, physics, and intelligent behavior.

Our team has raised $50M in seed funding from NVIDIA Ventures, Threshold Ventures, AIX Ventures, and notable angels including Naval Ravikant and Jeff Dean to build the foundational layer for the future of AI—powering everything from robotics training and simulation to digital twins and interactive environments.

We are looking for exceptional engineers to help build the simulation systems that will power the next generation of robotics and embodied intelligence.

The Role

We are hiring a Member of Technical Staff focused on Agentic Model Training to lead reinforcement learning infrastructure, model post-training, and the development of intelligent agents capable of operating across complex digital and physical environments.

You will work closely with Qi and the research team to improve large vision-language, multimodal, and code-generating agents through fine-tuning, reinforcement learning, trajectory data, reward design, and scalable evaluation.

Moonlake already has deep expertise in 3D and world-building. This role adds the model-training experience needed to systematically improve agent reasoning, planning, tool use, code generation, and long-horizon task execution while preparing the company for larger-scale reinforcement learning across interactive worlds, simulation, and embodied AI.

We are looking for a full-stack researcher and engineer who understands the complete model-training lifecycle, from data collection and training infrastructure to reward design, evaluation, and deployment. You should have strong research judgment, understand when training is necessary, which methods are likely to work, and what not to pursue.

What You’ll Do

• Build reinforcement learning and post-training pipelines for multimodal, vision-language, and code-generating agents

• Develop infrastructure for supervised fine-tuning, preference optimization, reward modeling, and reinforcement learning

• Create scalable systems for collecting, filtering, replaying, and learning from agent trajectories and interaction data

• Design rewards, verifiers, and evaluation frameworks for complex, long-horizon agent tasks

• Improve agents' ability to reason, plan, write and execute code, use tools, recover from errors, and complete complex workflows

• Develop training methods that improve agent performance across interactive 3D environments, simulation, and embodied tasks

• Scale distributed training, high-throughput inference, and rollout generation across multi-GPU environments

• Build reliable systems for evaluating model performance, tracking training improvements, and identifying failure modes

• Improve training reliability, reproducibility, observability, and computational efficiency

• Work across models, data, environments, rewards, and infrastructure to drive measurable improvements in agent capabilities

• Help define Moonlake's long-term strategy for agentic model training, robotics, and embodied intelligence

What We’re Looking For

• Real-world experience training large language, vision-language, multimodal, or code-generating models

• Strong experience in reinforcement learning, model post-training, or large-scale fine-tuning

• Experience building distributed training infrastructure or high-throughput inference and rollout systems

• Familiarity with supervised fine-tuning, preference optimization, reward modeling, and agentic reinforcement learning

• Experience with code-generation agents, tool-using systems, or long-horizon agent evaluation

• Strong Python skills and experience with PyTorch, JAX, or similar machine learning frameworks

• Understanding of model training across data collection, trajectory generation, reward design, optimization, and evaluation

• Ability to work across research and engineering to translate new training methods into reliable, scalable systems

• Strong research judgment, technical independence, and a bias toward building and shipping reliable systems

Preferred Experience

• Experience at a frontier AI lab or organization operating large-scale model-training systems

• Experience with code-model post-training, autonomous coding agents, or tool-using language models

• Experience training multimodal models, vision-language agents, or models operating in interactive environments

• Experience with robotics, simulation, world models, or embodied AI

• Experience designing verifiable rewards, outcome-based training systems, or long-horizon agent evaluations

• Experience scaling reinforcement learning workloads and distributed training across large GPU clusters

• Familiarity with trajectory-based learning, synthetic data generation, and agent-environment interaction loops

What Success Looks Like

Within your first year, you will have:

• Built Moonlake's core agentic model-training, post-training, and reinforcement learning infrastructure

• Created scalable systems for collecting and learning from agent trajectories across interactive environments

• Delivered measurable improvements in agent reasoning, code generation, tool use, and long-horizon task completion

• Established reliable reward systems, evaluation frameworks, and training pipelines that continuously improve model performance

• Helped the team determine which problems require training, which methods produce meaningful improvements, and which approaches should not be pursued

• Established a scalable foundation for larger-scale agentic reinforcement learning and embodied-model training

Why This Role Matters

Moonlake's agents must do more than generate content. They must understand complex requests, reason across vision and language, write and execute code, operate tools, build interactive worlds, and recover from mistakes.

Achieving this requires more than powerful foundation models. It requires training systems that allow agents to learn from experience, improve through interaction, and reliably execute increasingly complex tasks across digital and physical environments.

This role will build the foundational training infrastructure that enables Moonlake's agents to continuously improve, unlocking more capable interactive world models and the next generation of embodied intelligence.

We are committed to being an on-site, in-person team currently based in San Francisco.

We are committed to being an on-site, in-person team currently based in San Francisco.