My degree happens in the daytime. My life gets built after midnight. These are both true, and the second one would sound like a motivational poster if the first one didn't keep handing me the receipts.
I'm a naval architecture student. The days are spoken for: lectures, labs, the commute-shaped gaps between them, group work that expands to fill whatever space remains. By the time the ev...
In the last article, I said I keep rebuilding the same app. Before writing any code, I wanted to answer a fair question: don't tools for this already exist?
They do, and they're good. So here's what I found.
Starting a project today usually means picking from one of two families.
Yonto Player talks to a media server, or to a folder, through one interface: ContentSourceAdapter. The rule itself is covered in the opening post of the series. This one is about what sits behind the interface, and how it changed.
The short version is that the source contract started out as an HTTP API, moved into the app as JavaScript plugins, and is now the way every Jell...
Guardrails para advocacia são mecanismos técnicos e éticos que restringem o uso de IA em atividades jurídicas sensíveis — como redação de petições, análise de provas ou orientação processual — garantindo conformidade com o Estatuto da OAB e a Resolução 05/2023 do Conselho Federal da OAB. Eles impedem a autonomia plena da IA na prática da advocacia, exigindo su...
Guardrails para Claude, GPT e modelos próprios são camadas técnicas e políticas que limitam saídas inseguras, não éticas ou ilegais — aplicadas via filtragem de entrada/saída, classificação de conteúdo, regras de negação e RLHF. Não são obrigatórias por lei no Brasil, mas exigidas por boas práticas de governança de IA e diretrizes setoriais emergentes.
We’ve all been there. You spend hours meticulously engineering a prompt. You add a few "few-shot" examples, inject some XML tags to keep the output formatted, and wrap it in a clean Python script using the OpenAI SDK. It works brilliantly for the exact test case you wrote. Then, you hand it to a user, they type something slightly unexpected, the model hallucinates, your regex parser crashes on ...