Claude Code: 5 Production AI Agents for Insurance Claims

Описание: With this AI Agent Project, I show you how to use Claude Code to build a Production Insurance Claims AI agent team.

Free AI Agents Training + 2 Guides:
https://www.maryammiradi.com/free-ai-...
AI Agents Mastery: 7 Modules, 16 Projects, 60% OFF:
https://www.maryammiradi.com/ai-agent...

To reflect real-world complexity, we skip the toy chatbots and use real car damage images, repair estimates, policies, and claim history to build a 5-agent architecture: Intake, Damage Evidence, Policy, Risk, and Payout.

I also test the entire system with 72 evals, package everything inside a Gradio dashboard, and show you how to turn the whole build into a reusable Claude Skill and Claude Code plugin with hooks.

About Me:
I'm Maryam, PhD, built 400+ Production AI Agents, with 20+ years experience in AI, building and teaching production agents across 130+ countries.

🔔Subscribe for more Production AI Agents:
   / @maryammiradi  

Let's build together.

If you are building AI agents for the real world, this video shows the exact difference between using Claude Code as a basic coding assistant versus a controlled AI engineering workflow.

The 5-Agent Architecture:

This 5-agent system uses real car damage images, repair estimates, policy documents, claim history, LangGraph routing, Gradio, evals, audit trails, Claude Skills, and Claude Code plugins.

Intake Agent: Validates claim completeness
Damage Evidence Agent: Compares vision analysis against the customer story and repair estimates
Policy Agent: Verifies coverage, deductibles, and exclusions
Risk Agent: Scores anomalies and risk patterns
Payout Agent: Calculates final financial recommendations and routing

The system takes a claim, checks the damage image, compares it with the customer story and repair estimate, verifies the policy, scores the risk, recommends payout or human review, and logs the full decision trail.

I also show why evals matter. The system looked like it worked, but the evals caught a real production bug in the Risk Agent. After fixing it, 72 out of 72 tests passed.

Then I turn the whole build process into a reusable Claude Skill and Claude Code workflow.

What you will learn:

How to use Claude Code for production AI agent development
How to plan an AI agent system with BMAD
How to build a grounded insurance claims data pack
How to create a typed ClaimState with Pydantic
How to build a 5-agent insurance claims team
How to orchestrate agents with LangGraph
How to add human review routing
How to build evals for AI agents
How to create a Gradio dashboard
How to turn the build process into a Claude Skill
How Claude Code plugins and hooks can support production workflows

Tools and frameworks used:
Claude Code, LangGraph, Pydantic, Gradio, Pytest, BMAD, Claude Skills, Claude Code Plugins, Hooks

⏱️ Video Chapters:
00:00 - Building a 5-Agent Insurance Claims System
00:32 - Final Product Demo: Gradio Dashboard and Audit Trail
02:24 - Free AI Agents Training
02:41 - Architecture and Tech Stack: Claude Code, LangGraph, Pydantic
04:04 - Step 1: Claude Code Project Rules with CLAUDE.md
05:46 - Step 2: Architecture Planning with BMAD and BUILD_SPEC
06:59 - Step 3: Grounded Data with Kaggle Images and Synthetic Claims
09:17 - Why Data Is Most of the Real Work
10:16 - Pro Tip: Handling LLM Context Rot
11:24 - Expected Outcomes and Data Audit
12:18 - AI Agents Mastery 7-in-1
12:44 - Step 4: The Typed State Layer with ClaimState and Pydantic
13:40 - Step 5: Building the 5 Agents: Intake, Damage Evidence, Policy, Risk, Payout
17:39 - Step 6: LangGraph Routing and Audit Logs
19:16 - Step 7: Writing 72 Evals for System Testing
20:42 - Step 8: Building the Gradio Dashboard
22:24 - Step 9: Creating a Reusable Claude Skill
23:32 - Wrap Up and What's Next

#claudecode #aiagents #langgraph #claudeai #claude #aiengineer #aiengineering