Llm on Rafael Zago

How an AI works when you talk to it

Wed, 24 Dec 2025 00:00:00 -0300

Unpacking large language models

When you type a message to ChatGPT or Claude, what happens on the other side is simpler than it looks—but it runs at a scale that is hard to picture.

The essence: a digital football commentator

Picture this: World Cup final, Brazil vs Argentina, last minutes of the match. Galvão Bueno is calling the game, the ball reaches the Brazilian striker, he beats one defender, then another… and suddenly the audio cuts out. The broadcast keeps rolling, but the commentator’s voice is gone.

Skupper + InstructLab: controlling and protecting AI (act 1)

Fri, 02 May 2025 00:00:00 -0300

The story behind this solution pattern

Growing demand for AI-driven applications brings a hard problem: how do you deploy and operate AI models securely in environments that need strict data protection, while those models still have to be reachable by public services? That need became clear while building a local AI chatbot meant to handle sensitive and proprietary information — we needed a design that kept the model inside a protected environment.

InstructLab and Skupper: local AI without exposing data

Thu, 01 Aug 2024 00:00:00 -0300

Welcome to the Ollama Pilot.

Problem to solve

The main goal of this project is to create a secure connection between two sites, enabling the communication between the engineer machine and an Instruct Lab Model. The merlinite-7b-lab-Q4_K_M.gguf model will be used for the chatbot, and it is available in the Instruct Lab. The license of the model is available in the Instruct Labs.

But, why the banner? Well, the engineer needs to know who is better, Lebron or Jordan. The chatbot will be responsible for answering this question. The chatbot will receive the user input and send it to the llama3 model. The response from the merlinite model will be sent back to the user.