Multi-Agent AI Chatbot

Overview

Designed and built a production multi-agent AI chatbot for the Dish Network advertisement operations team. The system allows ad ops staff to navigate and utilize a complex advertisement order processing system through natural language — eliminating the need to memorize system-specific workflows and reducing onboarding time for new team members.

Problem

The advertisement order processing system had a steep learning curve. Ad operations staff needed to understand complex workflows, specific field names, and processing rules spread across multiple subsystems. Knowledge was siloed in individuals, and onboarding new team members was slow and error-prone.

Architecture

The system uses a multi-agent architecture orchestrated with LangGraph:

• Router Agent: Classifies user intent and routes to the appropriate specialist agent
• Query Agent: Retrieves order status, campaign details, and account information from the backend
• Action Agent: Executes order modifications, approvals, and status updates with human-in-the-loop confirmation
• Explainer Agent: Answers "how do I" questions using RAG over internal documentation

Each agent is powered by a foundation model via AWS Bedrock (Claude). LangChain handles tool calling, memory, and prompt management. LangGraph manages agent state transitions and conversation flow.

Key Technical Decisions

LangGraph over custom orchestration: LangGraph's stateful graph model made it straightforward to implement branching agent logic, error recovery, and human-in-the-loop checkpoints without building custom state machines.

AWS Bedrock for model serving: Leveraged Bedrock's managed Claude API to avoid infrastructure overhead and ensure enterprise-grade security and compliance.

Tool-calling with structured outputs: Each agent uses structured Pydantic models for tool inputs/outputs, ensuring type safety and predictable behavior when interacting with order processing APIs.

Results

• Deployed to production serving the full advertisement operations team
• Reduced average task completion time for common workflows by an estimated 60%
• Eliminated the need for system-specific training for 5+ common ad ops workflows
• Built using Generative AI-Driven Development practices adopted across the engineering team

Tech Stack

AI/ML:    LangChain, LangGraph, AWS Bedrock (Claude), Amazon Q
Backend:  Python, FastAPI, Databricks
Infra:    AWS Lambda, API Gateway, S3, IAM, VPC
Tooling:  Pydantic, pytest, GitLab CI/CD