ML Systems Research Engineer, RL / Inference / Evaluation

AMD AMD · Semiconductors · Santa Clara, CA · Engineering

ML Systems Research Engineer focused on building RL, inference, and evaluation infrastructure for AI-for-engineering systems. The role involves developing systems for agents and models to improve engineering workflows, including orchestration, evaluation, reward management, and feedback loops for model improvement. It emphasizes scalable ML systems for practical, repeatable research and production engineering.

What you'd actually do

  1. Build RL and inference systems for agentic engineering workflows, including job orchestration, sampling, scoring, caching, experiment tracking, and reproducible evaluation.
  2. Develop infrastructure for long-horizon and high-latency reward tasks where validation can take minutes to hours.
  3. Design staged rewards, proxy graders, sliced evaluation paths, retry strategies, and uncertainty-aware evaluation methods.
  4. Support optimization workflows with systems for candidate generation, benchmark execution, correctness checking, profiler feedback, reward modeling, and model-level improvement.
  5. Partner with AI research scientists on reward hacking research, reward shaping, metareasoning, and post-training methods for engineering tasks.

Skills

Required

  • Python
  • ML frameworks (PyTorch, JAX, TensorFlow)
  • ML systems
  • RL infrastructure
  • Inference services
  • Agent frameworks
  • Evaluation platforms
  • Distributed experimentation systems
  • Model inference
  • Batching
  • Sampling
  • Latency
  • Throughput
  • Observability
  • Reliability tradeoffs
  • Experiment design
  • Evaluation pipelines
  • Clear metrics
  • Logs
  • Reproducibility
  • Statistical discipline
  • Collaboration skills

Nice to have

  • Reinforcement learning
  • RLHF
  • GRPO
  • Preference optimization
  • Reward modeling
  • Reward shaping
  • Post-training systems
  • LLM agents
  • Tool-use systems
  • Code generation
  • Automated program repair
  • Compiler optimization
  • Benchmark-driven development
  • Distributed systems
  • Job orchestration
  • Kubernetes
  • Ray
  • Slurm
  • Workflow engines
  • Data pipelines
  • Large-scale experiment management
  • GPU systems
  • ROCm/HIP
  • CUDA
  • Profiling
  • Kernel benchmarking
  • Model serving
  • Distributed training/inference
  • Hardware engineering workflows
  • Design
  • Verification
  • Firmware
  • Simulation
  • Performance analysis
  • Publications or shipped systems in ML systems, RL, inference optimization, AI infrastructure, or hardware/software co-design

What the JD emphasized

  • scalable ML systems
  • long-horizon and high-latency reward tasks
  • reward hacking research
  • scalable inference and tool-use pipelines
  • Standardize datasets, eval definitions, run logs, leaderboards, failure taxonomies, and data collection for future training.
  • Reinforcement learning and post-training infrastructure for tool-using agents.
  • Inference systems for LLMs and agents, including latency, throughput, batching, sampling, reliability, and observability.
  • Evaluation systems for tasks with expensive, delayed, mixed, or sparse rewards.
  • Distributed experimentation, job orchestration, caching, data pipelines, dashboards, and reproducible run management.

Other signals

  • Reinforcement learning
  • Inference systems
  • Evaluation infrastructure
  • Agentic engineering workflows
  • Long-horizon rewards