AWS re:Invent 2025 - Real-time insights for smart manufacturing with AWS Serverless (CNS375)

Real-time Insights for Smart Manufacturing with AWS Serverless

Overview

This presentation discusses how AWS serverless architecture and generative AI can be used to build a connected manufacturing system with real-time insights and automated problem-solving capabilities. The key challenges addressed are data silos, skill gaps, and the inability to quickly respond to unplanned downtime events.

The Manufacturing Challenge

• Manufacturers often have data spread across multiple machines and databases, leading to disconnected data silos
• Experienced operators with deep domain knowledge are required to diagnose and resolve issues, but this knowledge is not easily shared
• Unplanned downtime events lead to significant financial losses, estimated at $1.4 trillion annually for the top 500 manufacturers globally

The Serverless Solution

1. Data Unification: Use AWS Glue to ingest data from various sources (IoT, documents, environmental sensors, etc.) into a central data lake
2. Integrated Data Model: Leverage the AWS Gluon framework to create a unified API and graph database, connecting data from disparate sources
3. Digital Twin: Build a digital replica of the production line to visualize real-time data and machine states
4. Generative AI Integration: Deploy a fine-tuned language model on Amazon Bedrock to enable natural language interaction and automated problem-solving

Key Components

1. Data Preparation:

   • Ingest data from various sources (IoT, documents, etc.) into an S3 data lake
   • Use a Step Function-based data pipeline to process the data and generate structured training data for the AI model

2. Model Fine-tuning:

   • Fine-tune a large language model using the structured training data, focusing on instructional and safety-oriented responses
   • Import the fine-tuned model into Amazon Bedrock for scalable, serverless inference

3. Intelligent Application:

   • Develop a web application with a conversational interface, leveraging the Bedrock API to interact with the custom model
   • Implement automated notifications and recommendations based on real-time data analysis and the fine-tuned model

Business Impact

• Enables a "connected factory" by unifying data from disparate sources and creating a digital twin of the production line
• Bridges the skill gap by providing operators with an AI-powered assistant that can diagnose issues, provide step-by-step instructions, and recommend corrective actions
• Reduces unplanned downtime and associated financial losses by enabling faster problem detection and resolution
• Facilitates continuous learning and model improvement through operator feedback and ongoing data ingestion

Example Use Case

• A cookie manufacturing line experiences an issue where half the cookies are burned
• The AI-powered system analyzes sensor data, machine logs, and standard operating procedures to determine that the oven temperature was uneven
• The system automatically notifies the operator, provides step-by-step instructions to adjust the oven settings, and recommends a maintenance check on the oven
• The operator follows the guidance, resolving the issue and preventing further production delays and waste

Resources

• AWS Gluon Framework GitHub Repo
• Implementing a Custom Model on Amazon Bedrock
• Building a Continuous Learning Pipeline