Senior AI / ML Engineer

The candidate will have responsibilities across the following functions:

Self-Hosted LLM Infrastructure:

• Deploy, fine-tune, and operate open-source models (Llama, Qwen, MedGemma, and successors) as our primary inference stack.
• Work with vLLM / SGLang / TensorRT-LLM for serving at scale, with disciplined attention to throughput, tail latency, batching, KV-cache, and GPU economics.
• Own fine-tuning workflows end-to-end (SFT, LoRA, QLoRA, DPO) on clinical notes, claims, and payer-rule data.
• Optimise GPU usage, latency, batching, and cost; make build-vs-buy and hosted-vs-self-hosted trade-offs explicit and measured.

Knowledge Graphs and Embedding-Based Retrieval:

• Design and maintain the knowledge graph encoding ICD-10-CM, CPT, HCPCS, modifiers, HCC, NCCI edits, LCD/NCD policies, and payer-specific rules and the relationships between them.
• Build embedding-based retrieval over clinical notes, historical claims, denial reasons, and payer-policy corpora, including chunking, embedding model selection, hybrid search, and reranking.
• Combine graph traversal and dense retrieval so every coded line, scrubbed edit, and appeal response is grounded in auditable evidence.
• Own ingestion, versioning, and quality of underlying knowledge sources (CMS, AHA, AMA, NCCI, payer bulletins).

Evaluation and Monitoring:

• Build continuous evaluation pipelines that gate every model, prompt, retrieval, and graph change before production.
• Run offline eval suites grounded in coder- and biller-validated labels; use LLM-as-judge where appropriate, calibrated against human ground truth.
• Monitor drift, hallucinations, regressions, and output quality in production; operate shadow-mode rollouts and per-cohort accuracy tracking (speciality, payer, chart type).
• Track business metrics: chart-level and opportunity-level coding accuracy, denial rate impact, clean-claim rate, cost per chart, and end-to-end latency.

LLM Systems and Prompt Engineering:

• Design prompts and context pipelines for coding (CPT, ICD, HCC, E/M), claim edits, denial classification, and appeal drafting.
• Implement structured outputs (JSON, function calling, constrained decoding) on top of the self-hosted stack.
• Apply RAG over medical coding standards (CMS, ICD-10 AHA, NCCI) and payer policies, grounded in the knowledge graph and embedding stores.
• Treat prompts as a thin, well-versioned, well-evaluated layer, never the load-bearing piece.

Agentic Workflows and Tooling MCP:

• Build MCP servers for internal tools: code lookup, NCCI / rule checks, payer logic, eligibility, denial classification.
• Design multi-step agent workflows with audit trails and human-in-the-loop checkpoints for coder, biller, and AR-analyst review.
• Define deterministic vs. LLM-based tool boundaries for reliability; reliability comes from knowing which is which.

Requirements:

• 5+ years in ML/AI engineering, including 6+ months in production LLM systems.
• Hands-on experience deploying and operating self-hosted LLMs (vLLM, SGLang, TensorRT-LLM, or equivalent).
• Strong experience designing embedding-based retrieval and/or knowledge graphs for grounded LLM applications.
• Demonstrated ownership of evaluation infrastructure, offline benchmarks, online monitoring, and drift and regression detection.
• Strong Python + PyTorch + Hugging Face experience.
• Production experience with monitoring, incidents, and system ownership.

Strongly Preferred:

• Fine-tuning experience (SFT, LoRA, QLoRA, DPO) on domain-specific corpora.
• Experience with graph databases (Neo4j, ArangoDB, or equivalent) and graph-aware retrieval.
• Experience with vector databases and hybrid search (BM25 + dense, rerankers).
• Familiarity with LLM observability tools (Langfuse, LangSmith, Arize, Braintrust, or in-house equivalents).
• Exposure to healthcare, RCM, claims, or other regulated domains.
• Experience with MCP or similar tool-orchestration frameworks.
• Strong prompt-engineering and LLM-evaluation instincts.