Agentic AI Engineer, Healthcare AI

About the Role

Deloitte is building AI-first reasoning models and agentic systems to fundamentally improve healthcare decision-making across payers, providers, and life sciences. This is an early, well-funded initiative focused on a ground-up rebuild of the decision-making machinery in American healthcare. You will own agent systems end-to-end, from architecture through production, with work shipping into live clinical and operational settings.

Key Responsibilities

Agent Architecture & Orchestration

• Design and implement agentic systems capable of multi-step reasoning, planning, tool use, and workflow execution against complex, regulated operational processes.
• Build stateful workflows using frameworks such as LangGraph and LangChain — including branching, retries, self-correction, human-in-the-loop checkpoints, and reusable orchestration patterns.
• Engineer for long-horizon reliability — multi-step task completion, recovery from compounding errors, planning under uncertainty, and robust tool use when individual steps fail.
• Build the reasoning behind regulated decisions — policy- and criteria-grounded outputs, structured proposer/critic/judge-style review, and auditable rationales for high-stakes decisions across the industry, from clinical review and prior authorization to claims integrity and care management.

Retrieval, Grounding & Context Engineering

• Develop end-to-end Retrieval-Augmented Generation (RAG) pipelines: ingestion, chunking, embeddings, vector and hybrid retrieval, reranking, contextual compression, and grounding strategies.
• Engineer memory and context management — conversational state, persistent memory, retrieval-aware context assembly, and token-efficient context selection.
• Apply modern context-delivery patterns (e.g., MCP-style tool/context interfaces) so agents access the right information at the right time.

Reliability, Evaluation & Safety

• Implement observability and tracing for prompts, tool calls, retrieval quality, agent traces, failures, drift, latency, and production behavior.
• Apply guardrails, safety controls, and failure-handling to reduce hallucinations and unsafe actions.
• Evaluate agents at the trajectory and task level — multi-step task success, failure-mode and regression analysis, and sandboxed test environments — alongside retrieval- and generation-quality metrics, automated checks, and human review.
• Engineer healthcare-grade safety — deployment eval gates, human-oversight and escalation models, auditability and traceability for regulated decisions, and PHI/HIPAA-aware data handling.

Integration & Production Craft

• Build integrations with internal and external tools, APIs, enterprise systems, databases, and model providers so agents operate safely within real business workflows.
• Deliver production-quality code with strong practices in testing, CI/CD, logging, versioning, and documentation; make architecture decisions that balance quality, safety, latency, cost, and model risk.
• Partner with modeling and post-training engineers to improve model behavior for tool use, grounding, and long-horizon reasoning — through evaluation-driven feedback and, where it helps, fine-tuned or reasoning-optimized models.
• Translate ambiguous, high-complexity operational processes into robust system logic and reusable AI patterns; stay current with advances in agentic systems and translate research into practical engineering decisions.

Requirements & Tech Stack

• Strong Python engineering skills and modern software practices: testing, CI/CD, version control, and API integration.
• Experience implementing observability, tracing, and debugging for LLM-based systems in production.
• Hands-on experience with at least one frontier model platform (e.g., Anthropic, Google, OpenAI) and/or open-weight/self-hosted models (e.g., Llama via vLLM), including production tool use and agent capabilities.
• Ability to travel 0-50%, on average.

Qualifications

• Bachelor's degree in Computer Science, Engineering, Data Science, Computational Linguistics, or a related field.
• Demonstrated depth building and shipping production agentic systems — this is your primary craft, not a recent exploration. We weigh shipped systems, research, model releases, and open source over years in a title; expect strong software/ML fundamentals plus substantial, recent hands-on agentic work.
• Strong, hands-on experience building production agent systems with modern orchestration — LangGraph/LangChain or equivalent, including custom orchestration.
• Experience designing and optimizing end-to-end RAG systems: indexing, retrieval, reranking, grounding, and evaluation.
• Strong understanding of memory and context management, including context windows, retrieval-driven context assembly, persistent memory, and high-signal context selection.
• Deep, practical understanding of LLM behavior — strengths, limitations, hallucination risks, reasoning constraints, and latency/cost trade-offs — and the evaluation methods used to measure them.
• Experience evaluating and debugging agent behavior — task-success and trajectory analysis, not just output quality.
• Familiarity with vector databases and retrieval infrastructure such as Pinecone, Weaviate, or Milvus (Preferred).
• Exposure to model adaptation and fine-tuning techniques such as LoRA or QLoRA (Preferred).
• Understanding of traditional NLP concepts: tokenization, semantic similarity, entity extraction, summarization, and transformer fundamentals (Preferred).
• Experience operating in highly regulated, high-stakes, or operationally complex environments; healthcare exposure — clinical, payer, or life-sciences workflows, or standards such as FHIR — is a plus, not a requirement (Preferred).
• Demonstrated habit of staying current with AI research, benchmarks, and emerging engineering patterns (Preferred).