Staff ML Data Scientist, LLM Recommender Systems We're looking for a top-tier ML Data Scientist to build, train, and deploy next-generation LLM-based recommender systems. You will work directly with a top notch science team, the AWS account team, and the AWS Applied Scientist leading the science. The role covers the full path from raw customer data to a deployed model. We are looking for a seasoned applied scientist with a multi-year record of production ML, not someone who recently moved over from analytics or data science generalism.
What You'll Do- Turn structured and unstructured traveler and catalog data into narrative text documents and synthetic data, at a scale of tens of millions of documents
- Make the feature and context engineering decisions: what information goes into each document and how to express it (e.g., market-relative price, tier, trip context)
- Run large-scale distributed training on Amazon SageMaker AI, including HyperPod
- Fine-tune models with LoRA/QLoRA adapters, supervised training, and reinforcement learning for ranking and reasoning
- Own deep learning modeling, including architecture, ranking heads, and co-training and multi-task setups
- Deploy low-latency, cost-efficient inference, including quantization, distillation, and profiling
- Build evaluation sets, run ablations, catch data leakage, and write up findings clearly for Expedia and the AWS science team
- Handle customer data safely, including personal-data removal, working inside the customer's environment
- Bring an AI-forward mindset to your daily work, using tools like Claude and Cursor to ship higher-quality work at pace
What You'll Bring- Spark (PySpark) and the Hadoop ecosystem (Hive, HDFS) on corpora of tens of millions of documents
- Hands-on Amazon SageMaker AI: processing, training, distributed training (HyperPod), and deployment
- LLM fine-tuning (LoRA/QLoRA) and reinforcement learning (e.g., GRPO or PPO-style methods)
- Strong deep learning fundamentals in PyTorch: architectures, training dynamics, co-training
- Large-scale batch and production LLM inference (e.g., vLLM, Amazon Bedrock batch)
- Data privacy and security on AWS: personal-data removal, IAM, KMS, Lake Formation
- Clear written communication with customer and science stakeholders
- Demonstrable usage of AI-forward tools such as Claude and Cursor
Helpful Extras and Unique Skills- AWS Trainium or Inferentia (Neuron SDK)
- Recommender systems: ranking metrics (NDCG, MRR), sequential and generative recommenders
- Embedding quantization or semantic IDs
- Customer-facing or embedded consulting experience
