AWS re:Invent 2025 - Streamline AI model development lifecycle with Amazon SageMaker AI (AIM364)

Streamlining AI Model Development Lifecycle with Amazon SageMaker

Overview

• Presentation on how to streamline the AI model development lifecycle and monitor/manage AI workflows using Amazon SageMaker
• Presented by Kushbush Shvastav (Senior Product Manager, Amazon SageMaker Studio), Bruno Piston (Senior Worldwide Specialist Solutions Architect, AWS), and Manikandan Pararmasan (Senior Staff Architect, COO)

The Growth of Generative AI

• Generative AI is transforming the business landscape rapidly
• IDC predicts global spending on generative AI will reach $22 billion by 2028, 32% of overall AI spending
• Goldman Sachs estimates generative AI can increase global GDP by 7% and lift productivity growth by 1.5 percentage points over 10 years
• 89% of enterprises are advancing generative AI initiatives, 92% plan to increase investments by 2027
• 78% of organizations use AI in at least one business function, 77% use AI models with 13 billion parameters or smaller

Challenges Faced by Enterprises

1. Disparate and disconnected ML tools significantly increase time-to-market
2. Isolation between team members (data scientists, AI developers, business teams) leads to duplicated efforts and missed opportunities
3. Governing AI/ML projects efficiently becomes exponentially complex at scale
4. Availability and management of infrastructure is key for training and fine-tuning ML/LLM models

How Amazon SageMaker Addresses These Challenges

• Amazon SageMaker Studio provides an end-to-end ML development platform
• Allows data scientists to build, deploy, and manage AI workflows in a single pane of glass
• Supports multiple IDE options (Jupyter Lab, Code Editor, R Studio)
• Enables data preparation, model selection, fine-tuning, and deployment
• Provides access to a hub of foundation models and fine-tuning techniques
• Allows running experiments, building pipelines, and monitoring models/endpoints

Key SageMaker AI Capabilities Demonstrated

1. Data Preparation:

   • Ability to prepare data interactively in Jupyter notebooks or using EMR/Spark at scale
   • Data can be stored in shared file systems (e.g., FSx for Lustre) accessible from SageMaker Studio

2. Model Training and Fine-tuning:

   • Support for various fine-tuning techniques (supervised, reinforcement learning)
   • Ability to leverage pre-built SageMaker Hyperparameter Tuning recipes for open-source models
   • Option to bring custom models and train them from scratch
   • Monitoring of training metrics and system performance using MLflow integration

3. Model Deployment:

   • Deploy models to production using SageMaker managed inference or on self-managed Hyperparameter Tuning clusters
   • Leverage pre-built LLM containers for quick deployment of generative AI models
   • Monitor deployed endpoints and perform offline model evaluation

SageMaker Studio Enhancements

1. Remote IDE Access: Ability to connect local IDE (e.g., Visual Studio Code) to SageMaker Studio compute resources
2. Trusted Identity Propagation: Propagate user identity across SageMaker workflows and downstream AWS services
3. Amazon Nova Customization: Customize Amazon's proprietary large language models (Nova Micro, Light, Pro) directly within SageMaker Studio

SageMaker Spaces: Accelerating Generative AI Development

• New SageMaker Hyperparameter Tuning add-on that enables running IDEs (Jupyter Lab, Code Editor) on the same cluster
• Provides a self-contained "space" with customizable compute, storage, and lifecycle configurations
• Allows AI developers to maximize cluster utilization through GPU sharing and fractional GPU support
• Enables unified governance and observability for administrators through SageMaker Hyperparameter Tuning task management

COO's Journey with SageMaker

• COO, a Canadian fintech company, faced challenges with vendor costs, performance requirements, security, and the need for an end-to-end ML platform
• Adopted SageMaker Studio as the core foundation for their in-house model development
• Leveraged SageMaker Studio for model development, SageMaker Pipelines for MLOps, and SageMaker Endpoints for real-time model serving
• Achieved 98% cost reduction ($1.47 million annual savings) compared to their previous vendor-based solution
• Maintained sub-50ms latency for their fraud detection use case, with improved accuracy and reduced false positives
• Expanded the use of SageMaker-powered solutions across various use cases, including loan underwriting, churn prediction, and generative AI/LLM applications

Key Takeaways

1. Amazon SageMaker Studio provides a comprehensive, end-to-end platform for streamlining the AI model development lifecycle.
2. SageMaker offers a suite of services and capabilities to address common challenges faced by enterprises, including disparate tools, team isolation, governance complexity, and infrastructure management.
3. The platform supports a wide range of ML/AI workflows, from data preparation and model training to deployment and monitoring, all within a single pane of glass.
4. Recent enhancements, such as remote IDE access, trusted identity propagation, and Amazon Nova customization, further improve the developer experience and enterprise-readiness of the platform.
5. The new SageMaker Spaces add-on accelerates generative AI development by enabling IDE-based workloads on the same Hyperparameter Tuning clusters used for training and inference.
6. Real-world examples, like COO's success story, demonstrate the tangible business impact of adopting SageMaker, including significant cost savings, performance improvements, and the ability to scale AI/ML solutions across the organization.