From Hacking iPhones To Building The First AI-Run Company
TL;DR
Pedro Franceschi was already a working engineer at 12 because iPhone jailbreaking in Brazil got him discovered — after teaching himself C/C++ and English via Google at age 8 or 9, he found an early iPhone 3G jailbreak, got press attention under the handle “PH,” and landed a part-time mobile engineering job with his mom sitting in on the interview.
His teenage hacking turned into real money fast: a Siri hack annoyed Apple, and an iPad app made him $300,000 at 14 — Franceschi reverse-engineered Siri to work in Portuguese, built “Quasar” for iPad multitasking, sold it for $10 through Cydia, and said his mother initially suspected he was doing something illegal online.
Brex started because US banks wouldn’t give startups corporate cards even when they had millions in cash — after dropping out of Stanford with cofounder Henrique Dubugras and pivoting from VR in YC, they built Brex around underwriting based on cash balances, not revenue history, which became the wedge into a market where Pedro says 97% still runs on traditional corporate cards.
Franceschi’s core product belief is vertical integration: you can’t reinvent finance UX without owning the stack — he connects Brex and his earlier Brazilian payments company Pagar.me to the same idea, building from “the metal” up through acquiring, issuing, and rails rather than depending on legacy vendors.
He says the Capital One deal is about scale, not surrender, and frames it with one image: ‘don’t crush the butterfly’ — Brex sold for $5.15 billion after previously peaking at a $12.3 billion valuation, and Franceschi argues the company had already reset internally by repricing equity, cutting nearly 30% of staff, and launching a “Brex 3.0” turnaround before deciding Capital One’s distribution and balance sheet could accelerate adoption.
The most forward-looking part of the episode is Pedro’s obsession with ‘virtual employees’ powered by AI agents — he describes Brex experiments with OpenAI/Anthropic-era systems like a recruiter agent named Jim on Slack and email, plus his own agentic ‘CEO autopilot’ that triages Slack, email, docs, WhatsApp, and meeting notes, to the point his wife enforces ‘no AI Saturdays.’
The Breakdown
From grief in Rio to code as refuge
Pedro Franceschi says the biggest luck of his life was discovering what he loved at 8 or 9: computers. After losing his father to lymphoma at age 8, he gravitated toward coding partly because software was a world where “you control all the variables” when the outside world isn’t controllable. He taught himself C and C++ in Brazil through Google while learning English at the same time, driven less by ambition than by pure curiosity.
The 12-year-old iPhone jailbreaker with a fake age and a real job
His first notoriety came from finding one of the earliest ways to jailbreak the iPhone 3G so it could be unlocked outside AT&T. A Brazilian reporter tracked down the kid behind the handle “PH,” and Pedro admits he lied and said he was 14 because 12 sounded too young to be taken seriously. That attention led to a mobile app company hiring him part-time; his mother joined the lunch interview and essentially told the founders, if anything happens to him, “I’m going to kill you.”
Hacking Siri, building Quasar, and accidentally making $300K at 14
Pedro’s teenage projects got much bolder: he reverse-engineered Siri to work in Portuguese, which triggered legal heat from Apple, and built Quasar, an iPad app that let multiple apps run like windows years before Apple supported it. He sold it for $10 in Cydia’s parallel app ecosystem and says he made about $300,000. The memorable part is his framing: at that age, the work had “no other reason besides love for the game.”
Why payments pulled him in, and why Brazil gave him an edge
Finance wasn’t the dream at first. He got dragged into payments through an iOS security job at a Square-like Brazilian company, then teamed up with Henrique Dubugras to build Pagar.me, effectively Stripe for Brazil, before they even knew Stripe existed. Pedro argues Brazil’s painful economics — including hyperinflation and the need for real-time settlement — made its rails more modern than America’s, while the US stayed stuck with ACH delays and fragmented banking.
Stanford, a failed VR detour, and the insight that became Brex
When Pedro and Henrique arrived at Stanford in 2016, they thought US fintech was solved and briefly tried to build in VR after playing with an Oculus demo at the Stanford Shopping Mall Microsoft Store. Within weeks, they realized they knew nothing about hardware or optics, pivoted in YC, and noticed startups with $7 million in the bank still couldn’t get a corporate card because banks underwrote on revenue history. That mismatch became Brex’s founding wedge: underwrite startups on cash.
The Capital One deal, the valuation reset, and Brex 3.0
Ashley presses on the awkward optics: Brex once hit a $12.3 billion valuation and sold for $5.15 billion. Pedro’s answer is unusually candid — in 2023 they faced reality, cut nearly 30% of the company, removed two management layers, repriced employee equity, and rebuilt around what he calls “Brex 3.0,” so the eventual outcome would still be a win for employees. He says what impressed him about Capital One was speed — roughly 40 days from idea to signed deal — and a founder-like philosophy from Rich Fairbank: “don’t crush the butterfly,” just “add water to the Brex plant.”
AI psychosis: recruiter agents, security wrappers, and a CEO autopilot
The back half turns into a live demo in words of Pedro’s agentic future. At Brex, he’s experimenting with “virtual employees” like a recruiter named Jim who lives in Slack and email; the interesting part, he says, is that Jim can build capabilities on the fly without anyone explicitly coding them. To make that safe, Brex built a monitoring layer called “crab trap” where one LLM watches another at the network layer — a kind of agent-on-agent supervision.
Pedro’s own life is already run by agents, down to movie tickets and follow-ups
Pedro says he now runs much of his job through OpenAI-style agent tooling, with a personal system called “lemon pie” that ingests Slack, email, WhatsApp, Google Docs, and Granola meeting notes, then maps everything to “people” and “programs” he cares about. It generates summaries, action items, and even drafts responses or executes tasks — including buying movie tickets near friends’ seats and using Brex virtual cards to pay online. The funniest human detail is his wife’s rule: “no AI Saturdays,” because otherwise he’d spend dinner dictating voice notes to his agents.