Agentic AI Engineer

You will build agentic AI systems that work on real clinical and life sciences data, multi-agent workflows, retrieval over messy domain data, and the evaluation and monitoring that make them trustworthy enough to ship. This is a build-and-own role. You take systems from prototype to production yourself.

Responsibilities:

• Build multi-agent workflows using LangGraph or similar routing, state, tool use, and human-in-the-loop review.
• Build retrieval systems over domain data: chunking, hybrid search, reranking, and knowledge-graph or ontology-based retrieval where it helps.
• Evaluate before you ship: test harnesses, LLM-as-judge, and regression suites, and monitor quality once they are live.
• Handle regulated data properly: de-identification, PHI controls, audit trails, and guardrails on inputs and outputs.
• Deploy and run what you build: containers, CI/CD, cloud, and safe rollback.
• Work directly with clinical and domain experts to turn unclear workflows into something a system can actually do.

Requirements:

• Strong Python and production API experience (FastAPI or similar).
• At least one agentic system you built that real users depend on; you can explain the design, what broke, and what you would change.
• Real RAG engineering, not just a vector store call. Retrieval quality work with numbers behind it.
• Working knowledge of an LLM evaluation and observability stack (LangSmith, Langfuse, Arize, RAGAS, or equivalent).
• Docker plus one cloud platform (AWS, Azure, or GCP), and comfort owning a service end-to-end. Judgment about when an agent is the wrong answer and a simple pipeline is the right one.

Preferred, not required:

• Life sciences or healthcare exposure is a plus, not a gate. We will ramp a strong engineer into the domain. We will not ramp a weak engineer into the craft.
• Clinical data, EHR/FHIR, pharma R& D, trials, regulatory, medical coding, or real-world data.
• Knowledge graphs / GraphRAG (Neo4j or similar) and ontologies such as SNOMED, ICD, or UMLS.
• Fine-tuning (LoRA/QLoRA) and knowing when it beats prompting or retrieval.
• Kubernetes, Terraform, MLflow.
• MCP or agent-to-agent protocols.