About the role 

We're building the systems that turn one of the world's largest scientific corpora into research intelligence. That means production NLP pipelines running over millions of journal articles, extracting entities, classifications, claim tuples, and summaries optimized for use by downstream agentic applications. We're looking for a senior data scientist to own domain-specific content modeling work end to end, from the eval set through the pipeline stage that ships it. 

You'll join a small, senior team where data scientists own their models in production. You'll write the code, own the evaluations, ship the changes, and stay accountable for the outcomes. This is a hands-on role for someone who wants to see their models through to real users in a rapidly evolving market. 

What you'll do 

• Design and build NLP enrichment pipelines that extract entities, classifications, claims, and summaries from scientific full-text at scale. 

• Compare NLP approaches to extraction and enrichment against LLM-based approaches, and pick the right tool for each task. That means putting traditional NLP (NER, sequence labeling, classification), embedding-based retrieval, LLM prompting, and fine-tuned smaller models on the same table, and defending each choice with evaluation, cost, and operational tradeoffs. This is a core part of the job, not an occasional exercise. 

• Own evaluation. Build the golden sets in consultation with SMEs and vendors, choose the metrics, and make productive tradeoffs between speed, quality, and cost. 

• Contribute to agentic AI application work: tool-using systems that reason over the enriched corpus, where your NLP and evaluation background will shape how the agent grounds and defends its answers. 

• Work directly with editors, product managers, and engineers. Bring the modeling perspective into product decisions, and translate stakeholder pushback into concrete modeling work. 

What you'll bring 

• Strong NLP background across modern (LLMs, transformers, embeddings, retrieval) and classical (NER, classification, sequence labeling) approaches. You've built evaluations and learned from the results. 

• Clean python.  You are comfortable in exploratory notebooks and production repositories, and an engineer taking over a modeling output from you has a good head start. 

• A habit of comparing approaches and choosing the right one for the task. You can defend "prompt a large LLM" and "train a small classifier on 2,000 labels" with equal seriousness, back the choice with an eval and a cost estimate, and know what to do when performance drifts. 

Nice to have 

• Experience working with scientific or scholarly text. 

• Familiarity with AWS (S3, Batch, Lambda, SageMaker) and Parquet or Iceberg data lake patterns. 

• Experience running LLMs under real cost and latency budgets in production. 

• Some exposure to agentic AI applications: tool use, multi-step reasoning, guardrails, and evaluation of trajectories rather than single-turn outputs. 

Why us 

We publish some of the world's most-read research, and we're now in a rare position: applying modern AI to a corpus of trusted scientific knowledge that spans two centuries. Researchers will use the systems you build here to move faster and get closer to the answers they came for. That's the work: from knowledge to impact.