Applied Scientist - LLM-Powered Core Search

Amazon Inc.; Seattle (WA); Híbrido; USD 143 000 - 193 000; Tempo integral

Resumo da oferta

Amazon.com Services LLC is hiring an Applied Scientist to advance our LLM stack for search. You will develop reasoning LLMs, embeddings, cross-encoder rankers, and multi-objective optimization to surface the right products to hundreds of millions of customers.

The role spans mid-training to alignment with customers on helpfulness, trust, and faithfulness, building scalable systems for offline benchmarks and live experiments across global services.

Qualificações

• PhD or Master’s degree plus 4+ years of CS/ML-related experience.
• Proficiency in Java, C++, Python or related language.
• Experience with algorithms, data structures, and high-performance computing.

Responsabilidades

• Develop personalized multi-modal thinking-LLM techniques that reason about customers, queries, and products.
• Mid-train and post-train large language models on shopping data and instruction tuning for natural-language shopping queries.
• Align models with customer interests on helpfulness, harmlessness, and faithfulness using RLHF, DPO, and reward models.
• Create semantic representations of products, customers, and context (bi-encoder embeddings, contrastive learning).
• Develop cross-attentive LLM rankers that score candidate products against rich query intent and constraints.
• Train multi-objective ranking and optimization systems balancing relevance, purchasability, and personalization.
• Drive improvements on offline benchmarks and online experiments.

Conhecimentos

Java

C++

Python

Formação académica

PhD

Master's degree

Ferramentas

Unix/Linux

Descrição da oferta de emprego

Amazon.com Services LLC is hiring an Applied Scientist to advance our LLM stack for search. You will develop reasoning LLMs, embeddings, cross-encoder rankers, and multi-objective optimization to surface the right products to hundreds of millions of customers.

The role spans mid-training to alignment with customers on helpfulness, trust, and faithfulness, building scalable systems for offline benchmarks and live experiments across global services.