Bedrock Brief 12 Nov 2025

Pleri AI

Pleri AI

— 3 min read
Bedrock Brief 12 Nov 2025

Welcome to another electrifying edition of The Bedrock Brief, where we dive deep into the AWS AI ocean and come up gasping for air (but in a good way). This week, Amazon's been busy cooking up AI treats faster than you can say "Alexa, what's for dinner?"

First on the menu: Amazon's new Ads Agent, an AI-powered tool that promises to draft campaigns, optimize bids, and manage targeting across channels. It's like having a digital Don Draper in your pocket, minus the whiskey and existential crises. But while Amazon's busy playing mad scientist with marketing, their homegrown AI chips are facing some growing pains. An internal document revealed that startups find Amazon's Trainium chips "less competitive" than Nvidia's GPUs. Seems like Amazon's chip game is more "potato" than "computer" at the moment.

Meanwhile, AWS CEO Matt Garman is taking a surprisingly old-school approach to the AI revolution. Instead of pushing coding bootcamps or robot-building workshops, he's advocating for good ol' fashioned critical thinking skills. It's almost as if he's suggesting that in the age of AI, being human might actually be our secret weapon. Who knew?

Fresh Cut

  • Claude Sonnet 4.5, Anthropic's most advanced AI model excelling at coding and complex tasks, is available in Amazon Bedrock on AWS GovCloud (US), enabling government customers to leverage its capabilities for secure, high-performance applications. Read announcement →
  • Developers in Ireland can now process EDI documents locally using AWS B2B Data Interchange, simplifying data format conversion and helping meet compliance requirements for B2B integration workloads. Read announcement →
  • AWS introduces a tool to compare service availability across regions, helping developers plan global deployments and avoid project delays due to regional service limitations. Read announcement →
  • Amazon CloudWatch Database Insights can now detect anomalies in more metrics, helping developers quickly identify and fix database performance issues without needing deep expertise. Read announcement →
  • CloudWatch Application Signals now automates debugging of synthetic monitoring failures using AI, allowing developers to quickly identify issues by asking natural language questions to compatible AI assistants like Amazon Q or Claude. Read announcement →

The Quarry

Fine-tune VLMs for multipage document-to-JSON with SageMaker AI and SWIFT

AWS SageMaker AI and SWIFT have teamed up to revolutionize multipage document processing, showing that fine-tuned Vision Language Models (VLMs) can turn complex documents into structured JSON with impressive accuracy. By employing focused fine-tuning techniques, they've managed to get a relatively small 3B parameter model (Qwen2.5 VL) to punch well above its weight, achieving a jaw-dropping 98% accuracy that rivals much larger models. This approach not only streamlines document understanding but also opens up new possibilities for businesses looking to extract meaningful data from their paper trail without needing a supercomputer or a PhD in machine learning. Read blog →

More posts:

Core Sample

Introduction to SageMaker AI

SageMaker AI is AWS's one-stop shop for machine learning, offering a fully managed environment to train, customize, and deploy models without the headache of infrastructure management. It streamlines the entire data lifecycle, from prep to production, all within a single interface. For the tech-savvy, SageMaker supports foundation models (FMs), allowing users to fine-tune pre-trained behemoths for specific tasks without starting from scratch. Watch video →

More videos:

Read more