Multi-step agent routing patient queries across billing, appointments, records, and emergency escalation

Status: Built & Tested (30+ test queries)

Tech Stack: LangGraph, GPT-4, OpenAI Embeddings, ChromaDB, Python

Github link - https://github.com/deepthiai8518/healthcare-support-agent

• Domain: Healthcare customer support
• Problem: Multi-intent queries + unsafe AI responses
• Solution: Agentic RAG with escalation as a terminal state
• Focus: Trust, failure modes, diagnosability

📌 Problem & Context

Business Problem

Healthcare support teams manually triage thousands of patient queries across billing, appointments, medical records, and emergency requests. Current systems fails due to :

• Semantic variations ("Why was I charged?" vs. "Explain my bill")
• Multi-intent queries ("Show my appointments AND related billing")
• Critical emergency detection buried in routine requests

Why AI?

The goal was to reduce manual triage effort while preventing unsafe automated responses in high-risk scenarios.

Traditional rules-based routing breaks down with:

• 50+ ways to phrase billing questions
• Context-dependent intent (e.g., "I need help" = appointment booking vs. medical emergency)
• Multi-step information retrieval across siloed systems (billing DB, appointment API, records vector store)