On the fourth day of the marathon toward R$1 million in email sales, the scoreboard was stuck: R$1,974 billed, 19 sales on autopilot. Instead of attacking the number directly, Guilherme decided to spend the whole stream on invisible infrastructure: a system so he'd never again find out about an AI development through a clickbait YouTube video. That's where the day's goal was born — an AI intelligence radar that, after a live brainstorm with the audience, got named "Sentinel."
The build started as usual: six VS Code windows open in parallel, each running Claude Code, and a long brainstorm inside his "brain" (the project's logbook) to design the architecture before writing a single line. They settled on a cron running every 30 minutes sweeping official blogs (Anthropic, OpenAI, Google, Meta, Mistral), GitHub Trending, Hugging Face, YouTube, and X, with one golden rule: total silence if nothing relevant shows up — no alert-for-alert's-sake. Triage stayed in two layers to save cost: Haiku filters fast, Sonnet only digs into what passes the filter, with Opus orchestrating everything via Cloud Flow. For YouTube, instead of counting views, they built an "outlier" criterion: a small video with disproportionate engagement scores higher than a big, lukewarm channel.
With the data model defined in Supabase (sources, items, YouTube channels), the cockpit gained a "Sentinel" tab showing sources, AI costs, and categories in real time. Guilherme tested the pipeline by sending an HTML email with a screenshot auto-generated by the deploy skill, and sent an AI-narrated audio message to the WhatsApp group explaining the project to the base of managers: the thesis was that the biggest problem for email marketers isn't writing the email, it's knowing what to write about. Meanwhile, Cloud Flow ran a swarm of agents in parallel, each terminal pulling the "brain" to know exactly its slice of the task.
The tensest moment of the night wasn't technical, it was security: mid-stream, the agent itself leaked an API key in the terminal output — twice. "We're going to have to reset that key," "generate a new one," "hide it silently" — the reaction was immediate, rotate the credential and move on, remembering that days earlier the project had already suffered an attack from a bad actor who created over 24,000 fake sign-ups trying to bring the system down. Building in public has that price, and the antidote is a fast reflex, not perfection.
The stream closed on an almost meta moment: Sentinel, still under construction, had already found something of real value — the launch of Anthropic's "Claude Partner Network" — and Guilherme applied to the partner program live, using the tool he was building to feed the tool itself. It also left behind a reflection that becomes the next day's task: the "brain" works fine as a logbook, but the project needs a living "system.md", a real architecture map synced with the code, so agents stop making decisions without full context.