AWS re:Invent 2025 - Building an AI-powered waste classification using Amazon Nova & IoT (AIM256)

Building an AI-powered Waste Classification System using Amazon Nova & IoT

Overview

This presentation showcases a solution developed by the AWS team in Santiago, Chile to address the challenge of proper waste sorting and recycling in their office. The system utilizes Amazon Nova, a multimodal AI model, and IoT technologies to create an automated waste classification system that guides users to deposit their waste in the correct containers.

The Problem and Motivation

• The AWS office in Santiago, Chile faced the common issue of employees and visitors struggling to properly sort their waste into the correct recycling, landfill, and compost containers.
• This problem is exacerbated in Latin America, where waste generation is projected to increase by 250% by 2050, while recycling rates remain low.
• In Chile specifically, 17 million tons of waste are generated annually, with the majority ending up in landfills rather than being recycled.

The Solution

• The team developed a waste classification system that uses a Raspberry Pi, webcam, and button interface to capture images of waste items and classify them using Amazon Nova.
• When a user places an item in front of the camera and presses the button, the system analyzes the image, determines the waste type, and displays the classification result on a connected tablet.
• The system provides clear visual guidance to the user on which container the item should be deposited in, based on predefined recycling rules for the AWS office.

Technical Architecture

• The solution uses a hybrid architecture, with edge processing on the Raspberry Pi and cloud-based AI inference using Amazon Bedrock and Amazon Nova.
• IoT Greengrass is used to manage the edge devices and connect them securely to AWS services.
• AWS Step Functions orchestrates the workflow, with a Lambda function preprocessing the image and calling Amazon Bedrock to classify the waste.
• The classification results are stored in a DynamoDB table using AWS AppSync's GraphQL API, and the web interface is built using AWS Amplify.

Key Features and Benefits

• High Accuracy: The system has achieved nearly 95% accuracy in waste classification, a significant improvement over the previous manual sorting process.
• Cost-Effective: The hardware setup costs only $350, making it a relatively inexpensive solution compared to other waste management systems.
• Scalable and Customizable: The system can be easily scaled to other AWS offices and customized by updating the waste classification rules in the natural language prompt.
• Measurable Impact: The solution has increased the recycling efficiency at the Santiago office by 52% over a six-month evaluation period, as measured by the building landlord.
• Open-Source: The team plans to open-source the solution in the coming weeks, making it available to a wider audience.

Real-World Deployment and Expansion

• The system has been deployed not only in the Santiago office, but also at the AWS Experience event, where it successfully classified waste at three recycling stations with no contamination incidents.
• The team is now planning to deploy the system in other AWS offices, including Buenos Aires, Bogotá, São Paulo, Arlington (HQ2), and New York.
• Several customers have also expressed interest in deploying the system in their own facilities, demonstrating the broad applicability of the solution.

Conclusion

The waste classification system developed by the AWS team in Santiago, Chile showcases the power of combining IoT, computer vision, and multimodal AI to address real-world sustainability challenges. By providing a cost-effective, scalable, and highly accurate solution, the team has demonstrated the potential for AI-driven waste management to drive meaningful environmental impact.