Stormguard

Autonomous, self-healing supply chain platform built on AWS serverless. Mitigates supply chain disruptions using agentic AI with human-in-the-loop to keep operations stable and profitable.

Comment

Inspiration

Natural disasters expose fragile supply chains. I wanted to build a system that could reason, act, and recover autonomously - protecting human welfare while helping enterprises build resilient supply chains.

Hurricane Milton (October 2024) caused $50+ billion in losses and revealed the problem: manual supply chain coordination fails during disasters. Retailers face 300-400% demand surges while teams scramble with spreadsheets for 48 hours - but hurricanes hit in 48 hours. The US faces 15-20 major disasters annually. What if autonomous AI agents could coordinate decisions in 30 seconds instead of 48 hours?

What it does

StormGuard: six specialized AI agents powered by Amazon Bedrock Claude Sonnet 3.5 coordinate autonomously during supply chain disruptions.

  1. Demand Intelligence - Forecasts demand surges from disaster severity and sales patterns
  2. Inventory Optimizer - Identifies stores at stockout risk, calculates reorder quantities
  3. Procurement - Creates emergency purchase orders with vendor selection
  4. Price Stability - Prevents disaster price increases (0% markup enforcement)
  5. Risk & Compliance - Validates decisions, triggers human governance for high-stakes orders
  6. Orchestrator - Coordinates agents and synthesizes executive summaries

Key Innovation: Specialized agents with separation of duties. Procurement creates orders, Risk validates them, humans approve high-stakes decisions. Real supply chain teams at machine speed.

Results: 90%+ service levels (vs 60% baseline), <30 second decisions (vs 48+ hours), 0% price gouging, complete audit trail.

How I built it

AWS Stack:

  • Amazon Bedrock - Claude Sonnet 3.5 with specialized system prompts per agent
  • AWS Lambda - Python 3.12 serverless orchestration
  • Amazon S3 - CSV data for transparency
  • AWS SAM - One-click Infrastructure as Code deployment

Data Pipeline: Enterprise-scale synthetic retail data: 50 stores, 200 products, 1.4M transactions, 600K inventory items, Hurricane Milton event patterns (165% demand surge).

Agent Coordination: Sequential transformation pipeline - each agent builds on previous output:

S3 Data → Demand Forecast → Inventory Requirements → 
Procurement → Price Validation → Risk Assessment → Executive Summary

Prompt Engineering: Specialized Bedrock prompts per domain - Demand Agent does forecasting, cannot make procurement decisions. Procurement Agent creates orders, cannot approve them. Strict boundaries prevent overreach.

Human-in-the-Loop: Risk Agent detects high-stakes decisions (>$500K spend, policy violations), pauses system, presents crisis justification for executive approval/rejection.

Challenges I ran into

  1. Sequential Agent Dependencies - Six agents passing structured data cleanly required precise output formatting. Each agent's response must match the next agent's input requirements exactly.

  2. Bedrock Prompt Engineering - Preventing agents from overstepping expertise was hard. Demand Agent tried making procurement decisions. Solved with strict system prompts defining boundaries.

  3. Lambda Timeout Management - Six sequential Bedrock calls pushed Lambda's 5-minute limit. Optimized by reducing token counts and streamlining data transformation.

Accomplishments that I'm proud of

  1. True Multi-Agent Architecture - Sequential coordination where each agent builds on specialized output, not just parallel API calls
  2. Ethical AI Built-In - Anti-gouging agent enforces 0% price increases during crises
  3. Production-Ready - One-click deployment, enterprise-scale data, human governance included
  4. Measurable Impact - 90%+ service levels, <30s decisions, full cost/revenue tracking
  5. Complete Transparency - Every decision shows source data, reasoning, output for auditability

What I learned

  1. Multi-Agent > Single LLM - Specialized agents with verification are more reliable than one model doing everything
  2. Bedrock Claude Sonnet 3.5 Excellence - Exceptional at structured reasoning, data transformation, consistent JSON outputs
  3. AWS Serverless Speed - Lambda + Bedrock + S3 let me focus on agent logic, not infrastructure

What's next for StormGuard

3 months: Expand disaster types (wildfires, port closures, winter storms), real-time weather API integration, multi-region supply chains, Focus on Inspiring internal Enterprise-scale builds with AWS

6-12 months: Retail chain pilot, ERP integration (SAP/Oracle), mobile executive approvals, historical simulation training

12+ months: Multi-modal agents with Bedrock vision (satellite imagery), predictive inventory pre-positioning, cross-retailer collaboration platform, open-source agentic supply chain framework

Built With

  • amazon-bedrock-(claude-sonnet-3.5)
  • amazon-web-services
  • aws-cloudformation
  • aws-lambda
  • aws-sam
  • boto3
  • python-3.12
Share this project:

Updates