AWS re:Invent 2025 - Build AI your way with Amazon Nova customization (AIM382)
Source: Dev.to
Overview
AWS re:Invent 2025 – Build AI your way with Amazon Nova customization (AIM382)
In this session, AWS Solutions Architect Veda Raman, Amazon product manager Dan Sinnreich, and Terra Security CTO Gal Malachi discuss customizing Amazon Nova models for security and content moderation. The talk introduces the new Nova 2 family (Lite, Pro, Omni, Sonic) and explores four customization techniques: Retrieval‑Augmented Generation (RAG), supervised fine‑tuning, alignment, and continued pre‑training.
Dan demonstrates how AWS customized Nova models with LoRA adapters to enable sensitive content moderation for cybersecurity, law‑enforcement, and media use cases while preserving core safety. Gal shares how Terra Security fine‑tuned Nova Pro to build AI agents for penetration testing, solving the “Guardrail Paradox” with a custom Guardrail Checker that raised true‑positive detection from 80 % to 92 %.
The presentation emphasizes that customization with proprietary data creates durable competitive advantages beyond generic model capabilities.
Introduction: Customizing Nova Models for Security and Content Moderation
Welcome everyone. I’m Veda Raman, Solutions Architect for Gen AI at AWS. Joining me are Dan Sinnreich, Product Manager, Amazon AGI (Responsible AI Controls), and Gal Malachi, Co‑founder & CTO of Terra Security.
Quick audience poll
- How many have used Amazon Nova models?
- How many have customized Nova or any model?
The goal of this session is to show you why and how to customize Nova models for your specific security and moderation needs.
Introducing the Amazon Nova 2 Family
The Nova 2 family, launched recently, includes four variants:
| Model | Key Characteristics |
|---|---|
| Nova 2 Lite | Price‑performant hybrid‑reasoning model; excels at agentic and tool‑calling use cases. GA available. |
| Nova 2 Pro | Highly capable multimodal model; strong performance on complex tasks such as coding and agentic workflows. |
| Nova 2 Omni | Multimodal reasoning model that accepts text, image, video, audio, and speech as inputs; can output text, images, and speech. |
| Nova 2 Sonic | Speech‑to‑speech model. |
All models support a 1 million token context window.
Why Customization Matters: From Generic AI to Business‑Specific Solutions
- Domain specificity: Gartner predicts that by 2027 more than half of enterprise Gen AI models will be domain‑specific. Generic models alone won’t meet specialized needs.
- Competitive edge: Accuracy is now baseline; differentiation comes from how you tailor the model to your data, workflows, and brand voice.
- Embedding proprietary knowledge: Customization lets you capture unique IP, business processes, and brand tone, ensuring consistent, context‑aware responses.
Four primary customization techniques
- Retrieval‑Augmented Generation (RAG) – Augments the model with external knowledge bases at inference time.
- Supervised fine‑tuning – Trains the model on labeled examples to improve task‑specific performance.
- Alignment – Adjusts model behavior to adhere to safety, policy, or brand guidelines.
- Continued pre‑training – Extends the base model’s knowledge with domain‑specific corpora before fine‑tuning.
Use Cases Demonstrated
Sensitive Content Moderation (AWS)
- Approach: LoRA adapters applied to Nova 2 models.
- Outcome: Enabled fine‑grained moderation for cybersecurity, law‑enforcement, and media scenarios while preserving core safety mechanisms.
Agentic Penetration Testing (Terra Security)
- Approach: Fine‑tuned Nova Pro and built a custom Guardrail Checker.
- Result: True‑positive detection improved from 80 % to 92 %, addressing the “Guardrail Paradox.”
Takeaways
- Customizing Nova models with proprietary data creates a durable competitive advantage.
- The Nova 2 family offers flexible modalities (text, image, video, audio, speech) to suit diverse applications.
- Leveraging techniques like LoRA adapters, fine‑tuning, and alignment enables secure, domain‑specific AI that aligns with organizational policies and brand voice.
Video reference: AWS re:Invent 2025 – Build AI your way with Amazon Nova customization (AIM382)