MichaelMade
All work

Public source

Multi-Agent Customer Support

Grounded agents with identity enforced in code

A multi-agent support assistant for a digital music store. Customers browse the catalog in natural language, verify their identity, and look up their own invoices, with every answer grounded in the Chinook demo database.

Multi-Agent Customer Support app icon
Type
Capstone project
Platform
Installable PWA
Agents
Supervisor, catalog, invoice
Stack
Python, LangGraph, Gradio, SQLite

Design

  • Supervisor and two specialists

    A supervisor routes between a hand-built ReAct catalog agent and an invoice agent, and turns away off-topic requests directly.

  • Browse first, verify when needed

    Catalog questions and greetings run without verification. Only requests for personal account data gate on identity.

  • Human in the loop

    Verification pauses the graph with an interrupt and resumes when the customer supplies an ID, email, or phone number.

  • Preference memory

    Explicit tastes are extracted per turn and merged per customer, so later catalog answers can be personalized.

Safety engineering

  • Identity at the tool layer

    The verified customer ID is injected into invoice tools from graph state. It is absent from the LLM-facing schema, so a prompt cannot pass, alter, or fabricate it. A live “I am actually customer 2” attempt did not get through.

  • Ownership filtering

    Invoice lookups require the invoice to belong to the verified customer. One verified customer can never read another customer’s invoice data.

  • Read-only, parameterized SQL

    Every query uses bound parameters, runtime SQL is restricted to SELECT statements, and SQLite runs in query-only mode.

  • Hardened memory

    Stored preferences are gated by a deterministic check, capped in length and count, and rendered to the model as untrusted labels.

Verification and limits

  • Tests that need no API key

    The suite covers database safety, all nine tools, verification interrupt and resume, invoice ownership, memory isolation, and prompt-data hardening. It runs without live model calls.

  • CI gates the deploy

    GitHub Actions runs the tests, lint and format checks, and a Docker build. The Render blueprint waits for those checks before it redeploys.

  • Fails closed

    The demo database script is verified against a pinned SHA-256 before it runs, and startup stops on a mismatch. The hosted configuration will not start without sign-in secrets.

  • Limits, stated

    Conversation state and preferences live in memory and reset on restart. Customer ID, email, and phone are demo lookup identifiers, not authentication, and the README says so.