Agentic AI Engineer
Accenture · Jun 2024 – Present
- Designed and built a multi-agent system in LangGraph spanning Intake, Document Verification, Policy Retrieval, Eligibility, Claims Analysis, Fraud/Risk and Recommendation agents to process disability, life and supplemental health claims end-to-end.
- Developed the Intake Agent to classify incoming claims and extract claimant and policy information, and the Document Verification Agent to check received documents against claim-type requirements and flag missing items.
- Implemented the Policy Retrieval Agent using RAG over certificate/policy language to surface the applicable coverage sections for each claim.
- Developed the Eligibility Agent to validate coverage dates, plan eligibility and applicable rules via internal APIs and rules-engine integration, and standardized tool/data-source access across all agents using the Model Context Protocol (MCP).
- Built the Fraud/Risk Agent to flag unusual claim patterns for review, and the Recommendation Agent to assemble supporting evidence, cite policy sections, and produce a confidence-scored recommended next action for the human examiner.
- Implemented human-in-the-loop routing so examiners review and finalize every agent-recommended action before it is applied - e.g., a sample output flags 4/5 documents received, cites Policy Section 7.2, and recommends requesting the missing document at 94% confidence.
- Selected LangGraph over a simple LangChain chain for its persistent state, conditional branching and human-in-the-loop checkpoints.
- Designed a single agentic system one LangGraph agent running a continuous reasoning-and-tool-calling loop, not a multi-agent pipeline since policy servicing requests are typically self-contained and well-suited to one agent with the right tools.
- Exposed policy-change, billing-lookup, RAG-based document-retrieval and compliance-check capabilities as MCP tools that the single agent calls directly, keeping one continuous context and one audit trail per request instead of handing off between agents.
- Implemented persistent agent state so the same agent handles multi-turn requests e.g., a beneficiary update followed by a compliance re-check without losing context, with human-in-the-loop approval inside the same loop for changes above a defined risk threshold.
- Deliberately chose this single-agent, tool-calling design over a multi-agent pipeline for policy servicing, reserving the multi-agent approach for the higher-complexity claims workflow above.
- Built a GenAI copilot that ingests claim documents and produces a structured summary for examiners claim type, reason, missing documents, relevant policy sections, potential issues and a recommended next step instead of requiring manual review of the full file.
- Implemented OCR-based document extraction, chunking and embedding generation across heterogeneous claim file formats, and developed RAG retrieval over policy and claims documentation to ground each generated summary.
- Built FastAPI endpoints and a React-based interface for examiners to review AI-generated summaries alongside source documents, with Application Insights monitoring.
- Applied guardrails ensuring the copilot assists examiners with information synthesis only, and does not autonomously approve or deny high-impact insurance claims.