About The Role
We're building the AI/ML enablement backbone for Nonstop Health - a shared platform on which every AI initiative ships as a governed, production-grade service. The flagship work today is an agentic platform for healthcare claims (EOB parsing, claim substantiation against member policy + IRS/MERP rules) built on self-hosted LLMs/VLMs, a multi-agent framework, and a drag-and-drop Agent Studio - all under HIPAA / SOC 2 / ISO 27001 controls. You'll own hard parts of this platform and stand up new initiatives on top of the same foundation.
This is a builder role for someone equally comfortable reasoning about an agent's failure modes, tuning a retrieval pipeline, and making a service survive an RDS failover. You care that AI systems handling sensitive, money-moving decisions are correct, grounded, observable, and defensible - not just demo-able.
What You'll Do
- Design and ship agentic systems - dynamic tool-calling agents (LangGraph + structured Pydantic outputs) that reason at runtime over a governed tool catalog, with per-step verification, grounding checks, LLM-as-judge, human-in-the-loop, and bounded, auditable loops.Build the RAG layer - document ingestion ? chunking ? embeddings ? hybrid retrieval ? reranking, with strict per-member/per-client access scoping and injection/poisoning defenses.
- Operate self-hosted inference at scale - vLLM (chat/vision) and an embeddings/rerank service on EKS GPU nodes; optimize throughput (continuous batching, prefix caching, quantization, KV-cache tuning) and enforce fair-share concurrency across tenants.
- Own the MLOps / governance plane - offline eval harnesses and golden sets, quality scoring, canary/gray-release with auto-rollback, cost/budget governance, and end-to-end observability (OpenTelemetry traces + metrics, immutable audit).
- Make it production-safe - durable state, circuit breakers, retries, PHI-safe logging/redaction, RBAC, fail-closed defaults, and horizontal scalability on AWS (EKS, RDS/pgvector, SQS, S3/KMS, Cognito) via Terraform.
- Extend the platform to new domains - take a new AI initiative from problem framing to a governed, evaluated, deployed service on the shared foundation, and mentor engineers on doing the same.
- Raise the bar - clean, typed, tested code (the platform holds a ruff + mypy + full-test-green bar on every change); thoughtful design reviews; pragmatic trade-offs between autonomy and determinism for high-stakes tasks.