Sam Levan

Setups · one leader, the real machines

A Mac mini, a local Git, and a pile of paper that searches itself

Maxime Gaudin

Maxime Gaudin · MadMax

CTO at Gladia, the speech-AI company, since November 2025. Before that, CTO at MadKudu through its acquisition by HG Insights, co-founder and CTO of Matcha, and part of the first ten at Malt, which he helped scale to 200 people.

LinkedIn

His second brain holds everything: every email, Slack, WhatsApp, SMS, meeting, analysis, contact. Including the letter his partner Julia read when she proposed. He hosts it at home, for privacy and security.

Why he started building his second brain

Three reasons, in his words: survive a new CTO job without working 20-hour days; get out from under ten channels of information (one API over all of them took hundreds of messages a day down to 30, the brain took it to 10); and walk into work already in control of the day, the calendar, the important emails, even the coffee topic.

Maxime on why he built it, from the episode. 60 sec.

How it fits together

The setup, line by line

Machine
A Mac mini at home, unreachable from outside the house.
Storage
Giddy, a small Git service he wrote himself. A local GitHub: every document lives there, every change is kept.
Paper
Paper Trail, an app on top of Giddy. A pile of scanned paper, indexed and searchable, no folder hierarchy anymore.
Front door
Cursor. Indexed PDFs and markdown out of the box, hybrid search, agent-agnostic, integrated with Git. 20 USD a month, unchanged for over a year.
Models
OpenCode + OpenRouter: pay per token, every model on the planet, switches constantly. Opus for thinking, Gemini for speed and OCR. A local model is not possible yet at his level; "the transition has started" with the data and the apps.
Notifications
None. WhatsApp is the only channel allowed to ping him.
Skills
His own library, the manual for how he does things, shared by all his agents. He ignores other people's skill libraries (they describe what the agent already knows). Two off the shelf: "caveman" (fewer tokens) and "I have ADHD" (kills verbosity).
Brand
brand.md, built with the company designer: logos, fonts, colors, spacing, and what to kill (the Claude gradient, the cards, Open Sans).

Paper Trail, how the search works

The one part worth a deep dive: a pile of paper becomes something you can ask questions of, and no human ever files anything.

IN Pile in the scanner no filing, no folders A model reads each page text, date, company, amounts A git repo of PDFs and JSON every page, every version, backed up ASK "payslip MadKudu" the question Text search first the exact words A fast model splits it semantic: payslip · exact: MadKudu Filters, sort weighted fields

From episode 1 of Sit with an AI native, recorded 2026-08-26. Reviewed by Maxime before publishing. What he does with it every day: the morning podcast, the decision log, decks grilled against the brain.