Born 1987 in Soledade de Minas, Brazil. Dismantled household electronics before he could explain why. First machine: a Pentium MMX in '97 — writing Delphi on Windows 95. Mentored into Python 2.0 by an ex-Microsoft engineer in the early 2000s; never let go of the language since.
Trained as a lawyer. Graduated top of class. Never practiced — walked away, traveled the world for a year, learned Russian, came back different. Co-founded Portal IDEA with his mother, a retired educator: a de-bureaucratized education platform now ranked among the top 10 worldwide in its sector.
$ locale — pt_BR · en_US · fr_FR · ru_RU · zh_CN (loading...)
# still asking the same question: how do things work?
Experiments in flight. Networks overfitting on private archives. Scripts mistaking the shutter for a loss function. The process is the artifact.
In [24]: import torch, pandas as pd
In [25]: df = pd.read_parquet("./archive.parquet")
In [26]: model = HardCut.from_pretrained("murad/base")
In [27]: model.fit(
...: df, epochs=∞, lr=3e-4,
...: loss="silence", grad_clip=1.0
...: )
epoch 001 loss=0.9124 val=0.8801 · warming up
epoch 042 loss=0.2841 val=0.3120 · overfitting nicely
epoch 137 loss=0.0417 val=0.0402 · the face gives in
epoch 240 loss=0.0031 val=0.0029 · ^C — enough.
Last login: Mon Jul 20 2026 from 127.0.0.1 pablo@murad:~$ cat /etc/motd __ __ _ _ ____ _ ____ | \/ | | | | _ \ / \ | _ \ | |\/| | | | | |_) | / _ \ | | | | | | | | |_| | _ < / ___ \| |_| | |_| |_|\___/|_| \_\/_/ \_\____/ > you are inside a machine that dreams. > train. eval. repeat. pablo@murad:~$ whoami pablo — data scientist / hacker (sometimes films) pablo@murad:~$ gpg --fingerprint pablo