AI Engineer - RL Post-Training and Agentic Systems

Description

The company is building robot foundation models and the infrastructure to train and deploy them. You will join the early AI team and work directly with AI leadership and researchers.

The team is early stage and runs as a disciplined fast follower. It plans to commit capital as deployment reaches its inflection point, building on proven foundations rather than early bets.

The company is substantially internally funded and also draws top-tier investors, so funding is not expected to limit the work.

The mission is to close the gap between research and deployment by turning real-world robot experience into better models quickly and safely.

Core skills and tools are listed below. Items marked as a plus are optional.

Responsibilities

• Define the infrastructure used to post-train robot foundation models with large-scale reinforcement learning, then deploy them as agentic systems across the robot fleet.
• Own the full loop: distributed rollouts in simulation and on real robots, policy optimization, and serving models on cloud and on-prem infrastructure.
• Build scalable RL post-training infrastructure for multi-billion-parameter models. Coordinate distributed actors, learners and reward models across GPU clusters, simulation and the robot fleet.
• Own cloud and on-prem serving. Deliver low-latency, high-throughput inference for RL rollouts and production robots, including fast weight synchronization between training and inference.
• Build the tooling to orchestrate, evaluate and monitor agentic systems that combine models with planning, tool use and robot skills for long-horizon tasks.

Requirements

• Have at least 3 years of hands-on experience building and running large-scale RL or post-training systems for large models, including actor-learner architectures, asynchronous rollouts and off-policy correction.
• Have deep expertise in RL post-training: PPO, GRPO and related policy-optimization methods, reward modeling, and frameworks such as veRL, OpenRLHF, NeMo-RL, TRL or Ray RLlib.
• Have deep expertise in high-performance model serving: inference engines such as vLLM, SGLang or TensorRT-LLM, batching, KV-cache optimization, quantization, and serving platforms such as Ray Serve, Triton Inference Server or KServe.
• Have experience with distributed systems: Ray, Kubernetes and/or Slurm, fault-tolerant orchestration of heterogeneous workloads, and monitoring with Prometheus and Grafana.
• Work in Python, PyTorch and/or JAX, CUDA, NCCL, and distributed training with FSDP or Megatron-Core.
• Enjoy the empirical side of AI research and want to speed up research with strong tools, so the team can make breakthroughs faster together.
• Combine a startup's agility with rigorous engineering, as a curious builder with a research mindset who supports researchers with solid software engineering.
• A plus: experience building agentic systems and evaluation harnesses for long-horizon, multi-step tasks.
• A plus: robot learning in simulation and sim-to-real transfer with Isaac Sim/Isaac Lab, MuJoCo or domain randomization.
• A plus: edge and on-robot inference on NVIDIA Jetson, Orin or Thor using TensorRT or ONNX Runtime under real-time latency limits.
• A plus: robotics middleware and data formats such as ROS 2 and MCAP.

Conditions

This is a full-time, permanent position.

Equal opportunity applies to all candidates.