Insights

Codex usage in Brazil grew thirtyfold in eight months. Enterprise seats grew fivefold in twelve. The distance between those two numbers is the amount of source code leaving the company by a path nobody drew.

Claude now watermarks every text it generates, and the EU AI Act requires the same of anyone serving a self-hosted open-source model. This post explains the mechanism (SynthID-Text, tournament sampling) with diagrams, reproduces the scheme on a local model with measured numbers, and closes with the recommendation for anyone who has to mark their own output.

A practical guide to PII with LLMs: what the LGPD considers personal data, the four layers of protection, and a real measurement — Presidio out of the box does not detect the Brazilian CPF. With the recognizers that fix it and the latency cost.

Distilling an SLM for a Portuguese customer-interaction analysis pipeline. The three-instrument evaluation protocol, what the arithmetic hides, and why the saving isn't that the GPU is cheap.

SLMs and self-hosted models are the only way to turn a contractual guarantee about your data into a technical one. This post is about the break-even math, which tasks actually fit, and how that becomes one line of routing.

I measured LiteLLM overhead against a local upstream answering in 0.36 ms, isolating the Docker network hop first. About 10 ms at sequential p50. Here is the method, the number, and how to decide between self-hosted and OpenRouter.

LiteLLM and Langfuse, six services, a virtual key per team with a spend cap, routing by task alias and fallback across providers. The full stack to download and bring up with one command.

Before any frame reaches the model, it has to be decoded, color-converted and copied across memory. That path often costs more than the inference. I measured CPU and GPU decoding on the same machine, with a real YOLO at the end of the path, and show what to use in each case.

Two of Brazil's largest technology operations arrived at the same architecture without coordinating: a layer between applications and models. This post is about what that solved for them, and when the problem becomes yours.

For an agent pilot to get approved, it needs planning, an estimate and, of course, a demo. Live, preferably. That is the problem: the agent that would work in production is the one that gives the worst demo. An autopsy of 73 real incidents, the four traits of the survivors, and the three questions to know which side yours is on.

Nubank, Bradesco and Amazon published huge gains from AI code generation. METR measured experienced developers getting slower. The contradiction is only apparent: the bottleneck moved, and the winners were the ones who treated verification as a process, not a stage.

Everyone wrote an AI usage policy. Almost nobody can say what technically happens when someone ignores it. This article is about the three real paths source code takes out of a company — and what actually closes each one.

Headroom, RTK, and MemStack: three open-source layers that cut token burn when you run Claude Code the way it was meant to be used.

When cost or compliance matters, OpenCode is a credible terminal assistant: your keys, your provider, repo-level config for teams.

Generative AI: Democratization and Innovation in the Base Model Layer

Elevating business efficiency - from acquiring top-tier talent to streamlining project execution, discover how staff augmentation revolutionizes the dynamics of AI and full stack development.

AI in Healthcare: A Prescription for a Smarter, Healthier Tomorrow

Unlocking Success: Why Outsourcing AI Development Is Your Smartest Move in Today's Tech-Driven World

Unlocking Success: Your AI Strategy - Building a Defensible Moat for Tomorrow's Business World

Building Bridges, Uncovering Insights: Demystifying Data Engineers and Data Scientists

Unlocking AI's Potential: Bridging the Skills Gap with Accessible Tech

Unlocking the Power of AI to Safeguard Your Data and Keep Threats at Bay

Unlocking the Power of AI to Safeguard Your Data and Keep Threats at Bay

Unlocking the Power of AI to Safeguard Your Data and Keep Threats at Bay

Unlocking AI: When, Why, and How to Embrace Machine Learning in Your Business