+ AI Agents · Flagship
LangGraph Multi-Agent Customer Support
Multi-agent e-commerce support with LangGraph: intent routing, tool-calling specialists, policy RAG and human-approved refunds via interrupt/resume, streamed live over SSE.
16/16routing accuracy, scripted evaluation
Overview
A supervisor graph classifies each customer message and hands it to an Orders, Product/Policy or Refunds agent. Each agent calls Pydantic-validated tools on a SQLite store or a BM25 policy index. Refunds above a threshold pause the graph with LangGraph interrupt until a human reviewer approves or rejects them in the web console, which also shows a live trace of every agent and tool step. Guardrails, per-thread SQLite memory and a FastAPI SSE backend complete the system.
Highlights
- 100% routing accuracy (16/16 turns) and 100% task success (15/15 scenarios) in a scripted evaluation against gpt-oss-120b
- Human-in-the-loop refund approval with LangGraph interrupt and Command(resume), with idempotent side effects
- Live agent trace: SSE stream of tokens plus router, agent, tool and approval events
- Guardrails: pre-LLM moderation, per-customer and per-agent tool scoping, Pydantic validation, PII-redacted logs, step limits
- 71 offline tests with a scripted fake chat model, plus live Groq integration tests