Senior Data Scientist, Alexa For Shopping (Rufus)

We are building a agentic intelligence system that transforms unstructured, noisy customer-data into actionable intelligence for product analytics to guide evolution of Amazon Shopping CX's — surfacing metrics on demand and insights unprompted, without an analyst in the loop.
We are solving one of the hardest problems in the agent driven data intelligence space to isolate insights from noise. This role will own multi agent system orchestration and context management; self-improving agent layer that gets measurably better over time without human intervention and reliable signal extraction from unstructured data and proactive intelligence that detects what matters before anyone asks.
Our agentic system is in production. What we don't yet have is a system that evaluates its own output quality, identifies where it fails, and closes that feedback loop automatically.

Key job responsibilities
As Senior Data Scientist, you will own the multi agent orchestration and the self-improvement system end-to-end. You will also own designing the overall architecture to extract insights from unstructured data at scale. You will work directly with the principal engineer, influence the technical roadmap across the team, and partner with SDE's.

- This role requires operating independently on problems that are not well-defined or structured, identifying and framing research challenges across broad problem areas, and delivering end-to-end solutions that have significant impact on the product.
- Own the multi-agent topology (Planner → Worker → Reasoner → Loop Controller) — inter-agent communication protocols, and loop termination logic
- Design and manage the context window strategy across agents
- Own all system prompts, routing prompts, and chain-of-thought scaffolding across agents
- Define what "better" means across dimensions (factual grounding, hypothesis novelty, evidence completeness, reasoning coherence) without ground-truth labels at scale
- Design how eval signal propagates back into prompt updates and model routing decisions
- Own schema grounding, sparse vector indexing, and domain-scoped kNN queries
- Own embedding strategy, intent classification accuracy, and entity extraction quality