Overview In this role you will advance agentic systems that assist operational decision making and orchestration within Amazon's fulfillment network. You will build end-to-end systems using multi-agent coordination, tool use, memory, and action execution, deploying models in Amazon buildings. You will train and evaluate LLM-based agents, lead research projects, and mentor interns while publishing findings. You will shape adaptive decision-making capabilities and contribute to scaling intelligent systems that harmonize physical and digital workflows.
Compensation / Benefits- health insurance (medical, dental, vision)
- RSUs and sign-on payments
- 401(k) matching
- paid time off
- parenteral leave
- EAP and Mental Health Support
Responsibilities- Generate training and preference data for use cases (reasoning trajectories, tool traces)
- Model reward structures and optimize policies for LLMs (DPO, IPO, RLHF/RLAIF with PPO/GRPO, rejection sampling)
- Fine-tune on step-by-step trajectories and tool-use traces
- Apply Verbal Reinforcement Learning and Continual Learning
- Implement RL for LLMs including offline RL and off-policy evaluation
- Develop agentic memory/state management (episodic/semantic memory, vector search, RAG grounding)
- Design evaluation metrics for decision quality and scale LLM-based evaluations
Key requirements- 3+ years building models for business applications
- PhD, or Master's degree with 4+ years in CS, CE, ML or related field
- Proficiency in Java, C++, Python or related language
- Experience in algorithms and data structures, parsing, numerical optimization, data mining, parallel/distributed/high-performance computing
- problem-solving mindset
- cross-functional collaboration
- mentorship and leadership
- Unix/Linux experience
- LLM training and fine-tuning
- reinforcement learning for LLMs (DPO, PPO, RLHF/RLAIF)
