The lab
One card per build: what it is, what it proves, and where to see it. Most of them run by themselves on a Mac at home.
A 3D map that draws cybersecurity companies as castles. Seven walls each, every fact with a source.
What it provesA method can be public while the data stays private. Six walls are judged from public sources, one is measured by a script.
Made withPython, a 3D page, a site scanner, Grok for X signals
See the methodA morning scan of Reddit for real things people built with AI. The best 5 land on my phone at 07:52.
What it provesCheap judgment at scale works. A decision model reads about 200 posts a day and picks the 5, for under 2 cents a day. A gate drops posters who never answer anyone.
Made withPython, Reddit feeds, a decision model, Telegram
See the archiveWhat people built with AI, and whether it survived. Rebuilt every morning.
What it provesChecking back is the value. Anyone can list launches. Few go back to see what is still alive.
Made withPython, static HTML
Open the radarOne post every 3 days, written by AI from real builds, traced to the original source.
What it provesAn autopilot is only as good as its gate. One checker blocks banned words, missing sources and leftover placeholders. A rejected post never goes live.
Made withClaude, Python, a scheduled job
Read the blogA practical AI guide most evenings, in English, Hebrew and French.
What it provesA summary is not a guide. Each one has a problem, steps and a prompt to copy, with sources as links at the end.
Made withClaude, Python, a scheduled job
Open the guidesShort question pages with a direct answer up top, written so AI search engines can quote them.
What it provesWriting for answer engines is a format, not a trick: one question, one direct answer, then the detail.
Made withMarkdown, Python
See the pagesA Telegram bot. I say an idea out loud, it gets transcribed on my Mac, researched and filed.
What it provesVoice notes in Hebrew can be transcribed on a laptop for free. And no AI runs unless a new idea arrives.
Made withTelegram, a Cloudflare Worker, local speech to text, Claude
How the second brain worksA free nightly collector reads 6 sources. Every Sunday the 5 best land on my phone.
What it provesYou do not have to follow everything. Collect for free every night, spend on picking once a week.
Made withPython, feeds, Claude, Telegram
Tests with a model that writes no text and only decides: it picks the radar's morning 5, and sorted 100 emails in a dry run.
What it provesDeciding is cheaper than writing. 400 posts judged in 57 seconds for $0.016. In the email dry run nothing was changed, on purpose.
Made withJev by TypeSafe, Python
A skill that takes one idea from a podcast, traces it to where it really started, and credits that.
What it provesA podcast is where you find an idea, not the source of it. The skill leaves an open question instead of inventing my opinion.
Made withA Claude skill, a taste profile
Six AI agents that prepare posts for any brand: scout, researcher, writer, editor, designer, publisher.
What it provesNo agent ever posts. I click publish myself, and every correction I make becomes a rule for next time.
Made withClaude agents
Written down so they don't get lost. If one of them is your problem too, tell me.
Or want to learn to build it yourself. Both start with a sentence.