AWS re:Invent 2025 - Build AI your way with Amazon Nova customization (AIM382)

Customizing AI Models for Specialized Needs: AWS Nova Customization

Introduction to Nova Models

• AWS introduced the new Nova 2 family of models, including:
  • Nova 2: High-performance hybrid reasoning model for agentic and tool-calling use cases
  • Nova 2 Light: Generally available lightweight model
  • Nova 2 Pro: Highly capable multimodal model for complex tasks like coding and agentic use cases
  • Nova 2 Omni: Multimodal reasoning model that accepts text, image, video, and audio input
  • Nova 2 Sonic: Speech-to-text model
• These models offer a range of capabilities and performance characteristics to suit different needs.

The Need for Customization

• Gartner predicts that by 2027, over 50% of AI models used by enterprises will be domain-specific.
• Generic, one-size-fits-all models are insufficient for specialized business needs.
• Customization allows enterprises to:
  • Capture unique IP and workflows
  • Align model responses to brand voice
  • Ground the model in proprietary knowledge
  • Improve accuracy and safety in domain-specific scenarios
  • Gain durable differentiation beyond generic models

Customization Techniques for Nova Models

• Retrieval Augmented Generation (RAG): Uses in-context learning to ground responses in your own knowledge
• Supervised Fine-Tuning: Trains the model on specialized data and tasks
• Alignment: Tunes the model to have a specific tone or brand voice
• Continued Pre-Training: Further trains the model on niche domain data

Customization Platforms

• AWS Bedrock: Provides a managed way to customize Nova models
• Amazon SageMaker: Offers pre-built recipes for fine-tuning and continued pre-training
• Nova Forge: Allows deeper customization with access to model checkpoints, custom reward functions, and responsible AI toolkits

Customizing for Sensitive Content Moderation

• Challenges with generic content moderation guardrails:
  • Security and cyber testing use cases may generate content that is blocked
  • Law enforcement, media, and online platforms may need to process sensitive content
• Customization approach:
  • Uses LORA adapters to "unlearn" specific alignment dimensions while maintaining core safety
  • Leverages content classification to allow-list specific content types
  • Provides customized models through Amazon Bedrock for on-demand inference

Customizing for Agentic Penetration Testing

• Penetration testing is a manual, time-consuming process that doesn't scale
• Terra Security's approach:
  • Uses Nova Pro as the base model, customized with security-focused content moderation
  • Adds additional "guard" layers to prevent generation of destructive payloads
  • Employs a "guard checker" agent to decide whether to allow or block payloads
  • Leverages human-in-the-loop data collection and fine-tuning to continuously improve the model

Key Takeaways

• Customization is essential for enterprises to adapt generic AI models to their specific needs and workflows.
• AWS provides a range of customization techniques and platforms, including Bedrock, SageMaker, and Nova Forge, to enable domain-specific model development.
• Customization can address challenges with generic content moderation, enabling use cases in security, law enforcement, media, and online platforms.
• Agentic penetration testing can be powered by customized Nova models, with additional safeguards to ensure reliability and safety.
• Continuous fine-tuning and human-in-the-loop data collection are key to improving customized models over time.