Job Description

What is The Role

The Search Conversational Experiences team builds Elastic’s new conversational (agentic) platform that lets customers chat with their own data in Elasticsearch. We own the quality layer for RAG, agents and tools, retrieval/citations, streaming, memory, and—crucially—the evaluation signals that turn open-ended questions into grounded, reliable answers. As a  Data Scientist , you’ll be part of a cross-functional team (backend, DS, PM, UX) driving chat quality end-to-end: designing and running evaluation pipelines, improving prompts and tool behaviors, and turning measurements into product decisions that customers can feel.se

You’ll help tackle frontier problems—folding RAG and vector search into an agent’s knowledge base, dynamically enriching model context to boost groundedness, shaping agent routing and tool selection policies, lighting up agent-driven visualizations on top of Elasticsearch data, and exploring multimodality and reasoning strategies where they truly move the needle. This is an applied role: you will prototype, evaluate, and partner with engineers to ship.

What you will be doing

• Own well scoped pieces of the offline and online evaluation pipeline for agent workflows: retrieval coverage, reranking quality, reasoning traces, tool selection accuracy, citation integrity, and final answer helpfulness and faithfulness


• Calibrate and validate LLM-as-judge rubrics against human labels, track agreement with statistics, and add periodic checks to prevent drift


• Instrument agent runs with traces so you can localize errors to retrieval, reasoning, tool execution, or grounding, then contribute CI checks that block merges on regressions


• Translate evaluation readouts into product calls such as model choice, routing policy, tool gating thresholds, prompt and chunking updates, and agent customization for Elastic use cases


• Collaborate with backend engineers on contracts for ES|QL, citations, and telemetry schemas, and with PM and UX to land findings in shipped features


• Share outcomes through clear docs, notebooks, and PRs, and contribute utilities that make evaluation faster and more reproducible for the team

What you will bring

• 3 to 5 years in applied DS or ML with production ownership, including at least 1 to 2 years focused on evaluating LLM or agent workflows in shipped systems


• Proven experience designing and running stepwise evaluations for agent pipelines: retrieval coverage, reranking quality, reasoning traces, tool selection accuracy, citation grounding, and final answer helpfulness and faithfulness


• Golden set hygiene: stratified dataset design, leakage controls, reviewer guidelines, inter-rater checks, and versioned labels


• Fluent with offline IR metrics and guardrails: Recall at k, nDCG, MRR, groundedness or citation support, plus latency and cost tracking; can move from offline gains to online A or B tests


• Telemetry and traces for agent runs that localize failures to retrieval, reasoning, tool execution, or grounding; ability to add CI quality gates that block merges on regressions


• Practical Elasticsearch experience or a similar search system; ES|QL familiarity is a plus


• Strong written communication and async collaboration habits in a distributed team

Job Tags

Similar Jobs