PROJECTS · BUILT & BACKED

Things I build.

Companies, research and bets. Everything here I either built with my twin or backed because I liked the founder. Building is where I put most of my energy: going deep into the details, owning the whole loop, and compounding 1% every day.

01 / ACTIVE

NOW

2023—∞
01

Didit

CO-FOUNDER & CEO2023 · Identity infrastructure · YC W26

Giving trust back to the internet. Developers integrate identity verification, proof of personhood, AML screening and fraud prevention through one API and SDKs — with a free tier that removed the price barrier the incumbents built their business on. We verify millions of humans across 190+ countries, and the product loop is fast: ship, measure fraud signal, talk to customers, ship again. This is where most of my time goes — end to end across product, code, distribution and customers. Profitable, multi-million dollar ARR, growing ~20% MoM with a tiny, obsessed team. Backed by Y Combinator (W26).

02

Sante

CO-FOUNDER2023 · Health & anti-aging · Spain

One of the biggest anti-aging businesses in Spain. We started as an aesthetic clinic, a med spa with a single location in Barcelona, and grew it into longevity: anti-aging protocols, hormone and biomarker optimization, my personal health obsession turned into a company. More than 10,000 patients have already visited us, rated 4.9 stars on Google across 700+ reviews. Now we're opening a new clinic in Madrid and our own lab-testing facilities — blood extraction, biomarker panels, lab infrastructure we control end to end. My role here is co-founder and high-level strategy; the day-to-day machine runs without me while my hours go to Didit. Profitable, doing $5M+ in annual revenue.

02 / ARCHIVE

BEFORE

LEARNINGS
04

Gamium

CO-FOUNDER2021–2023 · Web3 & metaverse · 33 people at peak

Our first company. We started with GPU farms mining Ethereum, then went deep into web3: blockchain protocols, a token launch, many smart contracts — and a metaverse. We were building Gamium as a decentralized metaverse with avatars and digital identity years before it was obvious that identity was the interesting part. Never found product-market fit and failed. The most expensive and most useful education we ever got.

05

StyleGAN2 → TensorFlow 2

May 2020 · Open source · Generative AI

StyleGAN2 is NVIDIA's style-based GAN architecture — for years the state of the art for photorealistic image synthesis (the tech behind thispersondoesnotexist). NVIDIA's official code was TensorFlow 1.x only, so I migrated the full config-f architecture (generator + discriminator, custom CUDA ops, CPU support) to TensorFlow 2 when no TF2 port existed. 142★ and my entry point into generative AI, years before the LLM wave.

RANDOM WALK IN LATENT SPACE
06

mil — multiple-instance learning library

September 2020 · Open source · AI Research

Multiple-instance learning is a weakly-supervised ML framework where labels exist at the level of bags of instances, not individual samples — exactly the shape of my university research detecting Parkinson's dyskinesia from sensor data. Reviewing the literature I found there was no standard, maintained MIL library in any language, so I built one: pip-installable, sklearn/keras-compatible, covering the classic algorithms, bag representations, and evaluation tooling so researchers can run reproducible MIL experiments without reimplementing papers.

Diagram showing multiple-instance learning bags and instances
MULTIPLE-INSTANCE LEARNING · BAGS VS INSTANCES