My name is Jorge. On the web, I usually go by jvican.
These days, I am taking a sabbatical after nearly six years at Netflix. I am using the time to play with computers, follow my curiosity, and figure out what I want the next chapter of my work to be about.
At the moment, that means taking the modern LLM stack apart and putting it back together: tokenizers, training loops, distributed training, inference systems, and GPU kernels. I am implementing pieces from first principles and working my way through PyTorch, JAX, Triton, CUDA, NCCL, and the layers below them.
I am interested not only in making models run and train at scale, but in understanding the nitty-gritty of why they run the way they do: how compiler abstractions map onto low-level hardware, where performance bottlenecks exist, and how to build ML infrastructure that helps researchers and engineers move quickly without giving up correctness and acceleration. That last question has been a recurring theme throughout my career, from compilers and build systems to ML research infrastructure.
This involves a bit of soul-searching, and deep technical study. The technical direction has by now become clear; what I am still exploring is where to apply it... which domains and problems are important enough to deserve the next several years of work. I am using this time to understand where my background and knowledge can be most useful, and take some risks.
Before this, I built ML infrastructure and research tools at Netflix, working extensively in Python across notebook infrastructure and large-scale ML systems, including a real-time feature platform used across personalization. Earlier, I worked at the Scala Center on the Scala programming language, compilers, and developer tooling.
Most of the public work below predates this current chapter. It comes from a time when I worked more openly on Scala tooling, but it still reflects the problems I enjoy: compilers, build systems, performance, and tools that help other people do better work.
Blog
I'm gradually reworking this site into a place to share notes, experiments, and occasional write-ups from the projects I've been exploring over the past few months.
Selected Projects
These are OSS projects I've significantly contributed to and/or co-founded:
- scalacenter/bloop, a build-tool-agnostic fast Scala build server.
- BSP, a build server protocol.
- sbt/zinc, Scala's incremental compiler.
- sbt/sbt, the most popular Scala build tool.
- scalacenter/scalac-profiling,
a profiler for the
scalaccompiler.
Selected Posts
These are mostly old posts from the time I was doing more developer tooling in the open.
- June 2019. Integrate Bloop with Bazel and Pants
- June 2018. Speed up compilation times with scalac-profiling
- Oct 2018. Overload methods with extra parameter lists
- Oct 2017. How often do we change our sbt build?
Selected Talks
iA few talks from my build-tooling years.
- Design
Challenges of Bloop, a fast concurrent build server
Scala World 2019, link to slides - Build Server Protocol and new IDEAs
Collaboration with Justin Kaeser @ ScalaSphere 2018, Kraków - Meet Bloop and Get More Productive v2
Collaboration with Martin Duhem @ Scala Days 2018, New York - Meet Bloop and Get More Productive v1
Collaboration with Martin Duhem @ Scala Days 2018, Berlin