GenAI SupplyChain Copilot

Banner
Process Flow
Agentic Flow
AWS Tech Architecture

Inspiration

Supply chains today are highly digital yet painfully unstructured. Procurement teams still manage RFQs, supplier quotes, and logistics plans through emails, PDFs, and spreadsheets. Data lives in silos across systems like Salesforce, ERP, and Slack — making every sourcing cycle slow, manual, and reactive.

What it does SupplyChain Copilot automates the RFQ-to-Fulfillment lifecycle using coordinated GenAI agents hosted on Amazon Bedrock Strands Agents.

It acts as an intelligent digital coworker that: Reads unstructured supplier emails and attachments to extract structured RFQ data. Creates or updates RFQs in Salesforce and DynamoDB. Identifies and scores suppliers based on price, delivery lead time, and quality. Normalizes quotes and evaluates each supplier’s reputation & sustainability using Wikipedia + Tavily web data. Drafts negotiation or award emails and posts updates to Slack via AWS SES + Comprehend sentiment checks. Plans optimal shipping routes (Air / Sea / Road) using live data from MapBox, AirLabs, Open-Meteo, and Folium for visualization. A Supervisor Agent (Nova Pro) orchestrates all six sub-agents, monitors state, manages retries, and produces an end-to-end audit summary of every sourcing decision.

How we built it Framework: Amazon Bedrock AgentCore Runtime + Strands Agents SDK

Models Used: amazon.nova-pro-v1:0 – Supervisor orchestration & reasoning anthropic.claude-3-5-sonnet-20241022-v2:0 – Negotiation, analysis, summaries anthropic.claude-3-5-haiku-20241022-v1:0 – Extraction & normalization

AWS Services: DynamoDB, SES, Textract, Comprehend, S3, CloudWatch External APIs: Salesforce, Slack, MapBox, AirLabs, Tavily, Open-Meteo Implementation Highlights: Each agent was implemented as a Python Strands Agent with specific tools and prompt logic. Supervisor orchestrates 0 → 5 execution via master-agent-runtime-entrypoint.py. Environment security handled through .env bootstrap and AWS Secrets Manager. Audit logs and agent outputs persisted to scm_* DynamoDB tables. Route maps dynamically generated using Folium + MapBox.

Challenges we ran into Designing tool-driven orchestration across three different Bedrock LLMs while keeping responses deterministic.

Managing secure credentials for Salesforce, Slack, and AWS APIs within SageMaker notebooks. Handling asynchronous dependencies between agents (e.g., RFQ creation before supplier evaluation). Balancing latency vs. accuracy when orchestrating multiple Bedrock model calls and API integrations. Ensuring reproducible audit logs and observability through DynamoDB + CloudWatch

Accomplishments that we're proud of Built the first end-to-end GenAI Supply Chain workflow completely on AWS Bedrock + Strands Agents.

Achieved a 70 % reduction in RFQ cycle time compared to manual procurement flows. Demonstrated explainable decision-making with structured audit trails and sustainability scoring. Seamless human-in-the-loop via Slack and SES notifications. Proved that multi-agent reasoning can orchestrate real enterprise workflows safely and transparently.

What we learned Agentic AI = collaboration > automation. Each agent must have a distinct role, tools, and schema.

Nova Pro excels at orchestration reasoning; Sonnet provides structured analysis; Haiku delivers fast text extraction. Prompt engineering discipline and schema validation are key to stable orchestration. Explainability and logging are vital for stakeholder trust in GenAI-driven decisions. Building securely with environment variables and .gitignore hygiene prevents credential leaks in AI projects.

What's next for SupplyChain Copilot – Multi-Agent Orchestration Integrate Amazon Nova Act for autonomous supplier-site browsing and PO creation.

Add Amazon Q Business for natural-language Q&A over RFQ and shipment history. Extend with a Predictive Disruption Engine using historical weather + port data. Introduce Voice RFQs: speech → text → agent pipeline. Apply Guardrails for secure, compliant outbound communication

Built With

agentcore-runtime
amazon-bedrock
boto3
comprehend
folium
geopy
mapbox
open-meteo
pandas
slack-sdk
strands-agents
tavily
textract

Submitted to

AWS AI Agent Global Hackathon

Created by

I have contributed to the complete development of the solution, starting from defining the roles of each agent to implementing all six primary agents using the Strands framework. I designed and integrated various tools to enhance the solution’s capabilities. I implemented an orchestrator agent to coordinate the agents and facilitate decision-making.
My work also involved integrating multiple AWS services, including AWS SES, AWS Textract, AWS Comprehend, Amazon DynamoDB and Amazon S3, as well as external APIs such as Salesforce, Folium, Mapbox, Overpass, OpenMeteo, AirLab and Tavily based on the solution requirements.
Furthermore, I implemented a structured thinking process and explainability features within the agentic flow to ensure transparency in the decision-making of each agent. I also documented the solution thoroughly, including process flows and technical architecture.

Bhuvaneshwari Patil
I led the design and end-to-end development of SupplyChain Copilot, architecting its multi-agent orchestration framework on Amazon Bedrock using Strands Agents. I defined the six agent roles (from RFQ ingestion to logistics optimization) and implemented the Supervisor Agent logic that coordinates their workflows via Bedrock’s AgentCore Runtime. I developed the Python codebase (supply-chain-master-agent.py, master-agent-runtime-entrypoint.py), integrated Salesforce, DynamoDB, Slack, and SES APIs, and configured environment security using .env bootstrap and token isolation. I also built the Folium + MapBox visualization module for multi-modal routing and optimized prompt engineering for Claude 3.5 Sonnet, Haiku, and Nova Pro models. In short, I was responsible for the system architecture, coding, and AWS integrations — transforming a concept into a working, demonstrable GenAI orchestration use case.

Najeeb Khan
Bhajandeep Singh

Updates

Bhuvaneshwari Patil started this project — Oct 22, 2025 08:05 PM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.