AI/ML Agent Engineer (Health Care Experience Required)

Role: AI/ML Agent Engineer

Location: Remote

Duration: Long Term

The role

HIP One runs 35+ agents in production against live Medicare prior authorization volume. Those agents are not prompt chains. They are typed, versioned, evaluated graphs with deterministic guardrails, per-criterion citation chains, and a routing layer that sends every adverse determination to a named human clinician before anything leaves the system.

As the AI/ML Agent Engineering FDE, you own that layer. You design the agent graphs, build the retrieval over clinical ontologies and regulatory corpora that grounds them, write the evaluation harness that gates them into production, and enforce the invariants that make a decision defensible in a CMS audit. Then you sit with the customer's clinical reviewers and tune against real cases until the outcomes land.

This is an engineering seat, not a research seat. We care about what survives contact with production volume, an auditor, and a utilization-management nurse who has thirty seconds to decide whether to trust your output.

What you will own

• Agent graph design on Aether One . Multi-agent decomposition of clinical decision workflows. State machines, not conversation loops. Tool design, control flow, retry and fallback semantics, partial-failure behavior, and typed contracts between agents. LangGraph is our common shape; the reasoning behind the graph matters more than the framework.
• Retrieval over clinical and regulatory knowledge. RAG across NCDs, LCDs, payer medical policy, formulary criteria, and coding references (CPT, ICD-10, HCPCS, LOINC, RxNorm, SNOMED). Chunking and indexing strategies that preserve citation granularity, because a decision has to point at the exact criterion it turned on. Hybrid retrieval, reranking, and retrieval evaluated separately from generation.
• Per-criterion citation chains. Every determination trace to the specific policy language that produced it. This is an architectural requirement, not a feature. If the chain breaks, the decision is not defensible, and if it is not defensible, we do not ship it.
• Evaluation harness and quality gates. Gold-standard case sets built with clinician adjudication. Regression suites wired into CI. Agreement metrics against human reviewers, drift monitoring on live traffic, and a clear-eyed view of where LLM-as-judge is useful and where it quietly lies. Eval pass rate is a number we publish, not a number we claim.
• Deterministic guardrails and the no-auto-deny invariant. Zero auto-denials is enforced in architecture, not in a prompt and not in a config flag. You build the routing, the human-in-the-loop gates on adverse outcomes, and the assertions that fail the build when an invariant is violated.
• Model routing and the containment boundary. Frontier reasoning is rented and swap-capable; the knowledge overlay is owned and stays customer-side. You work across that boundary: Claude and other frontier models where they fit, open-weight local substrates where sovereign or air-gapped deployment demands it, and the routing logic that decides which handles what.
• Read paths into the Healthcare Brain. Knowledge packs, MCP servers, and Claude Skills are how agents and customer teams consume governed context instead of prompts and hope. You build and maintain those surfaces for the workflows you own.
• Customer-facing agent engineering. You will sit with utilization-management nurses, pharmacy operations teams, medical directors, and the customer's own engineers. You will watch your agent handle their real cases, take the feedback in the room, and tune. You stay embedded through the first 90 days of production.
• Patent-grade specification work. Genzeon Platforms files patents on the architectures we build. ADRs you author may become claim language. We will train you on this, but you have to be willing to write with that level of care.

What we need you to bring

• Production agentic systems, not demos. You have shipped multi-step agents that ran against real traffic and real failure modes. LangGraph, LangChain, DSPy, Temporal-backed orchestration, or something you built yourself. What we look for is whether you can explain why the graph is shaped the way it is and what happens when step four returns garbage.
• Strong production Python. FastAPI, async patterns, PostgreSQL, pgvector or ChromaDB, clean typed interfaces, real test discipline. Comfort with at least one strongly-typed language is a plus. Code quality matters more than framework familiarity.
• Retrieval engineering depth. You have built and debugged RAG in production over messy, structured, and semi-structured source material. You know why naive chunking destroys citation traceability, you have measured retrieval quality independently of answer quality, and you have opinions about reranking that came from data.
• Evaluation as a first-class discipline. You have built eval sets rather than eyeballing outputs. Offline and online evaluation, human agreement measurement, regression gating, and honest calibration of where automated judges break down. If you have run evals against expert-labeled ground truth, lead with that.
• Prompt and context engineering at production scale. Structured output, tool and function-call design, context budgeting, failure containment. Treated as engineering with tests and versions, not as copywriting.
• Healthcare regulatory literacy. CPT, ICD-10, and HCPCS fluency. NCD and LCD structure. Medical necessity criteria and how InterQual or MCG-equivalent criteria sets are actually applied. CMS-0057-F at minimum. PHI handling discipline that is instinct rather than a checklist. This is a hard requirement for this seat, not a preference.
• Claude as coding partner, fluently. Spec-first prompting, agent-driven refactors, code review by Claude as muscle memory, Claude-assisted eval and test generation, ADR drafting. We expect Claude visible in your day-to-day, not as a toy.
• Customer-facing maturity. You can sit with a chief medical officer, a UM nurse, and a payer's head of engineering in the same room and lead a productive conversation about how an agent will fit their workflow without losing any of them.
• 5+ years of production engineering experience, with at least 2 building LLM or ML systems that ran in production. Healthcare or regulated-systems depth strongly preferred. Exceptional candidates slightly under the bar with unusually strong production agent work will be considered.

Nice-to-haves

• Fine-tuning, LoRA, distillation, or preference optimization on clinical or regulatory corpora
• Open-weight inference in production: vLLM, Ollama, llama.cpp, quantization, and the tradeoffs that come with each
• Vision-language pipelines for chart mining, medical record digitization, or document extraction
• MCP server authoring, or Claude Skills built for a production workflow
• Sovereign, air-gapped, FedRAMP-aligned, or Azure Government / AWS GovCloud deployment experience
• FHIR R4, Da Vinci CRD/DTR/PAS, X12 278, or NCPDP SCRIPT familiarity
• Published open-source or research contribution in agents, retrieval, or evaluation
• Patent author or inventor credit on AI or decision-system architecture
• Anthropic Claude API or Claude Code at production scale

Common expectations for every Healthcare FDE

• Healthcare Brain Academy, Builders Track. Every Healthcare FDE works through Steps 1 to 4 of the Builders Track before customer engagement. CPT / ICD-10 / HCPCS literacy, regulatory framework, production discipline, and clinical validation are the floor, not optional reading.
• Claude-First engineering. Spec-driven development. ADRs within 24 hours of design decisions. Invariant assertions in every ADR. Claude is the coding partner across the SDLC; we expect comfort and discipline using it, not aversion to it.
• Ownership of the outcome. FDE means you own the customer outcome end-to-end, not the JIRA ticket. If the workflow does not work in production, the work is not done, regardless of how clean the code looks.
• Production-grade discipline. Healthcare AI in production has different gates than experimental ML. Per-decision citation chains, audit-grade explainability, cryptographic decision ledgers, and human-in-the-loop on every adverse outcome are not features we add later. They are architectural requirements from line one of code.
• Customer-facing communication. FDEs talk directly to customer engineering, clinical, compliance, and business stakeholders. We expect engineers who can write a defensible compliance memo as comfortably as they write an agent graph.

What you will work on in your first 90 days

• Day 1 to 14. Healthcare Brain Academy Builders Track. Deep-dive on the production agent graphs running in WISeR: the code, the eval suites, the guardrail layer, the open defects. Pair with the Practice Lead on an active engagement and read the ADRs and patent specs in scope.
• Day 15 to 45. Ship your first agent to a non-production environment behind a full eval gate. Likely a criterion evaluator on a specific NCD or LCD, or a retrieval improvement to an existing decision path. Full ADR, gold-standard case set, and CI-wired invariant assertions.
• Day 46 to 90. Production ship. Your work goes live behind a clinician-in-the-loop gate for non-affirmations, and you own the first month of production behavior: the drift, the edge cases, the tuning. You will be customer-facing by day 60 at the latest.