Perplexity serves tens of millions of users daily with reliable, high-quality answers grounded in an LLM-first search engine and our specialized data sources. We aim to use the latest models as they are released, but the intelligence frontier is a jagged one, and popular benchmarks do not effectively cover our use cases. In this role, you will build specialized evals to improve answer quality across Perplexity, covering search-based LLM answers and other scenarios popular with our users.
Responsibilities
Architect and maintain automated evaluation pipelines to assess answer quality across Perplexity's products, ensuring high standards for accuracy and helpfulness
Design evaluation sets and methods specifically to measure the impact of tool calls (particularly web search retrieval) on the final answer's quality
Develop VLM-based solutions to programmatically evaluate how final answers render visually across different platforms and devices
Continuously review public benchmarks and academic evaluations for their applicability to the Perplexity product, adapting and incorporating them into our regular performance measurements
Operate within a small, high-impact team where your evaluation metrics directly shape product changes, collaborating closely with technical leadership to measure and improve Answer Quality
Qualifications
PhD or MS in a technical field or equivalent experience
4+ years of experience in data science or machine learning
Strong proficiency in Python and SQL (expected to write production-grade code)
Experience building within a modern cloud data stack, specifically AWS and Databricks
Comfortable with agentic coding workflows and using AI-assisted development tools to iterate faster
Preferred Qualifications
1+ years of experience working with LLMs at scale, specifically with LLM-as-a-judge setups
Prior experience working on customer-facing web products or consumer apps, with real user traffic at scale
A strong research background, with experience applying research methods to real-world ML problems
Experience defining evaluation metrics (e.g., factual consistency, hallucination rate, retrieval precision) and building ground truth datasets