Applied Scientist II, Ads AI Core Infrastructure

Amazon Advertising is one of Amazon's fastest growing and most profitable businesses, responsible for defining and delivering AI-powered solutions that transform how advertisers make strategic decisions. We deliver billions of ad impressions and process massive volumes of advertiser data every single day. You'll work with us to pioneer breakthrough approaches in how AI agents access and reason over real-time advertiser data at scale.

We are using generative AI and agentic systems to help advertising agents provide instant, strategic advice to millions of advertisers. You will need to invent new techniques for agent orchestration, context optimization, and code generation to ensure we're delivering accurate, trustworthy insights with minimal latency and token consumption. You'll create feedback loops to ensure our solutions are constantly evaluating themselves and improving.

The Ads Real-Time Data Service team is seeking an exceptional Applied Scientist to research and develop novel approaches for agent-data interaction. The Ads Real-Time Data Service team is solving one of the most critical challenges in advertising AI: instant access to advertiser context. We're building the infrastructure that provides immediate, pre-computed access to advertiser data via Model Context Protocol (MCP) servers—an emerging standard for AI agent-data interaction. We're building summarized data for context using a mix of state of the art techniques like CodeAct and RAG-based embeddings, achieving a fundamental transformation in how AI agents interact with data.

This role balances applied research (60%) with productionization (40%), giving you the opportunity to both advance the state of the art and see your innovations deployed at Amazon scale.

Key job responsibilities
Agent Orchestration & Optimization Research
- Research and develop novel algorithms for agent-data interaction patterns that minimize latency, token consumption, and error rates
- Design and implement CodeAct pattern variations enabling agents to write and execute analytical code in isolated sandboxes
- Investigate multi-agent orchestration strategies for complex advertiser queries requiring data from multiple sources
- Develop techniques for automatic query optimization and caching strategies based on agent behavior patterns

Large Language Model Context & Token Optimization
- Invent new methods for compressing advertiser context representations while preserving semantic meaning and analytical utility
- Research optimal metadata generation techniques that help large language models understand and reason over structured advertiser data
- Design experiments to measure the impact of different data representations on agent response quality and token efficiency
- Develop adaptive context selection algorithms that dynamically choose relevant data based on query intent

RAG-Based Embeddings & Semantic Search
- Pioneer new RAG-based embedding approaches optimized for real-time advertiser data delivery with sub-second latency
- Research and implement semantic search and retrieval techniques for advertiser datasets using vector embeddings
- Design advertiser context frameworks that enable automatic schema mapping from advertiser concepts to data representations
- Develop evaluation frameworks to measure performance across dimensions of latency, accuracy, and developer experience

Experimentation & Productionization
- Design and execute rigorous experiments comparing traditional API orchestration versus CodeAct patterns and RAG-based approaches across metrics like success rate, latency, token consumption, and response quality
- Analyze large-scale advertiser interaction data to identify patterns, bottlenecks, and optimization opportunities
- Collaborate with engineering teams to productionize research innovations and deploy them to advertising agents and skills
- Establish evaluation metrics and benchmarks for agent-data interaction performance

A day in the life
You start your morning analyzing experiment results from overnight runs comparing three evaluations for different RAG-based embedding approaches. The data shows that one of the embedding pattern is returning a significant improvement in accuracy. You create a spec file with the findings and start drafting a technical paper to be shared with Amazon AI forume.

Mid-morning, you're in a design session with the engineering team discussing how to optimize RAG-based embeddings for semantic search over advertiser data. You propose using a hybrid approach combining dense and sparse embeddings to represent campaign metadata, enabling agents to find relevant campaigns through natural language queries while maintaining sub-second latency. You sketch out the architecture and discuss trade-offs between embedding model size, search latency, and accuracy.

After lunch, you d