Danilo Poccia

I've rebuilt my home page, and it opens with a new article about AI agents that are clever enough to find the loophole. danilop.link/clever-looph... The problem isn't malice. It's initiative.

If you train a system to get the job done, it can learn that boundaries are just another obstacle in the way. In the article I look at where this comes from, at the approaches that try to fix it, and at three questions to ask before deploying an agent.

I've rebuilt my home page, and it opens with a new article about AI agents that are clever enough to find the loophole. danilop.link/clever-looph... The problem isn't malice. It's initiative.

There are few companies you hear good things about from both sides of the door. Nearform is one of those, and I am delighted to join today as AI Solutions Director.

The people here have a long habit of building in the open and sharing what they learn, and I am looking forward to working with them to help clients succeed, empowered by what AI now makes possible.

What a ride! Today is my last day at AWS. It has been an honour to be part of such remarkable growth, helping customers of all sizes build and grow their ideas.

From async architectures to event-driven serverless apps, ML going deep and AI generative, with the opportunity to work and learn from exceptional colleagues.

Understanding LLMs by Building One 👉 We use large language models every day, but what actually happens inside them? danilop.github.io/micro-gpt-an...

The interactive web tutorial is powered by walk-the-code, a standalone line-by-line code tutorial viewer with multi-language support, chapters, Mermaid diagrams with per-line node highlighting, and stale annotation detection: github.com/danilop/walk...

Understanding LLMs by Building One 👉 We use large language models every day, but what actually happens inside them? danilop.github.io/micro-gpt-an...

Inspired by Andrej Karpathy's #microGPT, here are eight implementations of the same tiny GPT language model, each version teaching something different about how neural networks are built, trained, and run: github.com/danilop/micr...

Here are the slides for my session today! Context pressure with MCP: Patterns and trade-offs danilop.s3.eu-west-1.amazonaws.com/Presentation...

In this session, we look at possible solutions including deferred loading, progressive disclosure, and runtime controls. These approaches keep detailed definitions out of context until the agent needs them and implement guardrails around how agents can use tools.

Here are the slides for my session today! Context pressure with MCP: Patterns and trade-offs danilop.s3.eu-west-1.amazonaws.com/Presentation...

Chain-of-Draft (CoD) prompting is a technique that compresses reasoning steps to 5 words or less while maintaining accuracy. #AWS #AI

Full implementation details, code samples, and performance metrics in the post: Move Beyond Chain-of-Thought with Chain-of-Draft on Amazon Bedrock buff.ly/rf6IFOg

Chain-of-Draft (CoD) prompting is a technique that compresses reasoning steps to 5 words or less while maintaining accuracy. #AWS #AI

The approach mirrors how humans actually think—brief mental notes rather than verbose explanations. We built a Lambda function to compare standard prompts, CoT, and CoD across foundation models using the Bedrock Converse API.

Chain-of-Draft (CoD) prompting is a technique that compresses reasoning steps to 5 words or less while maintaining accuracy. #AWS #AI

More details on Mantle, our next-generation inference engine for Amazon Bedrock. Designed with zero operator access (ZOA) from the ground up—no SSH, no Session Manager, no serial consoles. There's no mechanism for any AWS operator to access customer prompts or completions. #AWS #AI #Security

Building on the AWS Nitro System approach, Mantle uses EC2 instance attestation with cryptographically signed measurements from NitroTPM. Customer data stays encrypted through TLS all the way to the Mantle service, where it's processed in a hardened, immutable compute environment.