Senior Data Scientist, AI Retrieval Systems

Roughly 25 to 30 million people in the United States live with a rare disease. There are somewhere between 7,000 and 10,000 distinct rare conditions, and the large majority have no FDA-approved treatment.

Research on these conditions keeps running into the same obstacles. Published evidence for any one disease is thin and scattered across sources. The same clinical finding gets written down a dozen different ways depending on who recorded it. And the people with the most at stake, patients and their families, are usually the least equipped to read the specialist literature written about their own condition.

Large language models are well suited to this class of problem, and the research programs we support are investing in applying them carefully. In this role you will build the retrieval and knowledge layer that those AI systems stand on. That means the disease and phenotype vocabularies that give a model something precise to reason over, the semantic search that finds the right concept behind an imprecise human phrase, and the ranking that decides what a user sees first. Ontologies serve as internal scaffolding throughout. Users should never have to see one or learn what it is.

This is a senior individual contributor position with unusual range. You will own the data layer, the retrieval services built on top of it, the interfaces where results become visible, and the path onto the computing infrastructure that runs it all. You will work directly with NIH program staff, clinical geneticists, and rare disease information specialists.

• →Model biomedical knowledge for rare disease research. Ingest disease and phenotype ontologies and controlled vocabularies into PostgreSQL with a maintainable release and refresh path, reconcile identifiers across sources, and work through term hierarchies to determine what is clinically relevant for a given condition.
• →Build retrieval-augmented services that ground everyday language in clinical concepts. Embed term labels, definitions, and synonyms, retrieve candidates, and have a model disambiguate against context before any value is committed.
• →Treat retrieval as a database problem. Tune keyword and vector search over large biomedical corpora, and be ready to defend the recall and latency trade-offs you choose.
• →Build the ranking and relevance layers that decide what surfaces first, including domain-aware weighting and graceful degradation when a condition falls outside curated coverage.
• →Deliver the interfaces where this work becomes visible to users, in Next.js, React, and TypeScript. This covers question and confirmation flows, result presentation, and live status for long-running pipelines.
• →Deploy continuously onto NIH on-premises and high-performance computing Kubernetes environments. Helm charts, StatefulSets, secrets, ingress, GPU scheduling for self-hosted inference, and scheduled jobs are all in scope, and you will partner with the operations teams that run those environments instead of standing up parallel cloud infrastructure.
• →Build the evaluation that tells us whether retrieval and concept mapping are good enough to rely on, and keep it running as a regression suite instead of a one-time measurement.
• →Log what the system does and why. Request identifiers, latency, errors, and which concept the system selected all need to be captured, so that staff can review an AI-assisted result instead of taking it on faith.
• →Work out what researchers, clinicians, and patient communities need, and turn it into data models, retrieval behavior, and interface design.
• →Write the work up. You will contribute to manuscripts, conference abstracts, and posters with NIH investigators, and you will be credited as an author on work you helped produce.

• Bachelor’s degree in Data Science, Computer Science, Bioinformatics, Biomedical Informatics, or a related field. An advanced degree is preferred. We will consider equivalent professional experience in place of a degree.
• At least 5 years building and operating production software or data systems. At least 2 of those years should involve shipping LLM-powered applications (agents, retrieval, or evaluation) that people depend on. We weigh depth in retrieval and applied LLM engineering more heavily than total years.
• Experience building retrieval systems end to end, covering indexing, query construction, and measuring retrieval quality against real data.
• Experience evaluating systems that have no single right answer, using golden sets, offline regression suites, or metrics such as Recall@K and MRR to decide whether a change was an improvement.
• Experience with structured output and tool or function calling, meaning you have constrained a model to a typed schema and validated what came back.
• Ability to own a service end to end, from schema design through deployment and operation.
• Ability to obtain and maintain a Public Trust Security clearance.

• Python, with FastAPI, Pydantic, and pytest.
• PostgreSQL at depth, covering vector search (pgvector or equivalent), full-text search, embedding pipelines, indexing, and query tuning.
• LLM application engineering: provider APIs and gateways, prompt and context design, structured generation, and tool use.
• Data ingestion and transformation pipelines with a repeatable refresh path.
• Containers and Kubernetes, enough to ship, debug, and operate a service on infrastructure you do not administer.
• Working comfort in Next.js, React, and TypeScript.
• Git-based collaboration and CI/CD in a shared codebase.

• Biomedical ontologies and controlled vocabularies, including MONDO, HPO, UMLS, MeSH, and other OBO Foundry resources, along with comfort working through term hierarchies, synonyms, and cross references.
• Grounding model output in a domain terminology through embedding-based retrieval plus model disambiguation, such as entity linking, concept normalization, or ontology alignment.
• Helm, and deployment to on-premises or HPC Kubernetes environments.
• Serving open-weight models in production with Ollama or vLLM behind a gateway such as LiteLLM, and work with domain embedding models such as MedCPT.
• Background in rare disease, clinical genetics, or translational research.
• Contributions to an open biomedical resource, standard, or consortium, such as OBO Foundry ontologies or GA4GH.'
• Published or presented work that explains your engineering to people who did not build it. Peer-reviewed papers, conference talks, preprints, technical blog posts, and public open source contributions all count.
• Prior or current NIH experience.

We are looking for an engineer first. If you have shipped retrieval systems that people depend on and have never opened an ontology file, we want to hear from you. Rare disease and ontology background is useful but not required, and we expect to teach the domain to whoever we hire. Candidates who meet the required qualifications and none of the preferred ones are encouraged to apply.

Disclaimer: The above description is meant to illustrate the general nature of work and level of effort being performed by individuals assigned to this position or job description. This is not restricted as a complete list of all skills, responsibilities, duties, and/or assignments required. Individuals may be required to perform duties outside of their position, job description or responsibilities as needed.

The diversity of Axle’s employees is a tremendous asset. We are firmly committed to providing equal opportunity in all aspects of employment and will not tolerate any illegal discrimination or harassment based on age, race, gender, religion, national origin, disability, marital status, covered veteran status, sexual orientation, status with respect to public assistance, and other characteristics protected under state, federal, or local law and to deter those who aid, abet, or induce discrimination or coerce others to discriminate.

Accessibility: If you need an accommodation as part of the employment process please contact: careers@axleinfo.com

This role has a market-competitive salary with an anticipated base compensation range listed below. Actual salaries will vary depending on a candidate’s experience, qualifications, skills, and location.