About the job
We're hiring an engineer to help us bring reinforcement learning to every agent team at NVIDIA. This is a rare chance to shape how autonomous, self-improving agents learn and evolve across the enterprise. The role sits at the intersection of ML research and production engineering. What if every agent developer could add self-improvement loops to their workflows without needing deep RL expertise? That's the challenge here: evaluate emerging approaches, adapt them into enterprise-ready blueprints, and make them available inside sandboxed execution environments with the security and governance the enterprise demands. We believe the best training and self-evolving agent platforms come from people with diverse backgrounds and want this person to help us build ours.
Responsibilities
The work splits between creating enterprise-ready RL capabilities and partnering with agent teams to put them into practice.
Building RL cookbooks and environments:
Evaluate and adapt democratized RL approaches into reusable cookbooks and blueprints so agent developers can integrate self-improvement loops (GRPO, DPO, PPO, RLAIF) on their own
Design verifiable reward environments building on NeMo Gym, extending to domain-specific environments for internal use cases
Operationalize NVIDIA and third-party training backends as production services inside Sandbox
Integrate with NeMo Microservices (Curator, Customizer, Evaluator, Guardrails) to enable end-to-end data flywheel workflows for RL
Infrastructure, reliability, and collaboration:
Lead data curation and active learning strategies to continuously improve training data quality
Design RL training loops for agent self-improvement: reward modeling, policy optimization, safety constraints
Integrate with AI Factory GPU infrastructure for throughput, data locality, and multi-node training
Build observability for training runs and ensure workloads meet security and governance requirements
Collaborate with platform, security, agent infrastructure, and internal customer teams on safe deployment of training outputs
Qualifications
Minimum
MS in CS, ML, or related field (or equivalent experience)
10+ years of experience
Experience operationalizing fine-tuning methods (LoRA, SFT) and especially RL techniques (DPO, GRPO, PPO, RLAIF) into reusable cookbooks and self-service workflows
Familiarity with distributed training frameworks (e.g., Megatron, NeMo, DeepSpeed, FSDP, HF Accelerate) and ML ops skills covering pipeline automation, job orchestration, and GPU cluster management are important here
Proficiency in Python, Go, Rust, or similar
Background in CS, ML, or related field through formal education or equivalent experience
Preferred
Building RL environments or training recipes that other teams consumed as self-service capabilities
Familiarity with NVIDIA infrastructure (DGX, AI Factory, NVLink/InfiniBand), NeMo Microservices, or the evolving RL-for-agents ecosystem (rLLM, Agent Lightning, HUD, OpenRLHF, SkyRL)
Experience with data curation, active learning, continuous learning loops, or data flywheel architectures also valued