AWS re:Invent 2025 - Edge AI in Real-World Solutions using AWS SageMaker, Greengrass and Bedrock

Summary of AWS re:Invent 2025 - Edge AI in Real-World Solutions using AWS SageMaker, Greengrass and Bedrock

Introduction to Edge AI

• Edge AI is the deployment of AI models on edge devices like smartphones, cars, drones, and PCs
• Key drivers for edge AI include:
  • Latency - Performing AI inference locally is faster than sending data to the cloud
  • Cost - Offloading compute to edge devices can reduce cloud compute costs
  • Privacy - Keeping data and processing on the device improves privacy
• Historically, deploying AI models on edge devices has been challenging due to the need to optimize for different hardware architectures

Streamlining the Edge AI Deployment Process

• Qualcomm has developed a set of tools and services to simplify the edge AI deployment process:
  1. Training in the Cloud: Use existing tools like AWS SageMaker to train AI models
  2. Model Optimization: Qualcomm's AIHub service automatically optimizes trained models for Qualcomm hardware
  3. Deployment Pipelines: Edge Impulse provides a platform to package optimized models and deploy them to edge devices

Detailed Deployment Workflow

1. Train Model in the Cloud: Use AWS SageMaker or other tools to train an AI model, outputting a standard format like ONNX or PyTorch.
2. Optimize Model for Edge: Send the trained model to Qualcomm AIHub, which automatically optimizes it for target Qualcomm hardware and provides performance metrics.
3. Prepare for Deployment: Use Edge Impulse to package the optimized model into a deployable artifact, including testing on sample data.
4. Deploy to Edge Devices: Leverage AWS IoT Greengrass to efficiently deploy the model package to a fleet of edge devices running Qualcomm processors.
5. Run Model on Edge: The Edge Impulse Linux Runner executes the optimized model on the edge device, providing low-latency inference results.
6. Monitor and Analyze: Send the inference results from the edge devices to AWS IoT Core, allowing further analysis and monitoring in the cloud.

Key Results and Benefits

• 20x faster model throughput by optimizing with Qualcomm AIHub
• Millisecond-level inference latency on edge devices
• Seamless deployment to edge device fleets using AWS IoT Greengrass
• Ability to perform real-time inference and analysis at the edge

Real-World Use Case: Pill Defect Detection

• The presented solution used the described workflow to build an edge AI application for detecting defects in pills on a factory floor conveyor belt.
• Key aspects:
  • Training a computer vision model in AWS SageMaker
  • Optimizing the model for Qualcomm hardware using AIHub
  • Packaging the model with Edge Impulse for deployment
  • Deploying to edge devices using AWS IoT Greengrass
  • Performing real-time defect detection at the edge
  • Sending results to the cloud for monitoring and analysis

Conclusion and Takeaways

• The combination of Qualcomm's edge AI tools and AWS cloud services provides a streamlined end-to-end workflow for deploying sophisticated AI applications at the edge.
• Key benefits include reduced development time, optimized model performance, efficient fleet deployment, and the ability to leverage both edge and cloud capabilities.
• The pill defect detection use case demonstrated the real-world applicability of this approach, highlighting the potential for edge AI to enable new classes of intelligent, low-latency applications.