Yogesh Sen

Who I am#

I'm Yogesh Sen. I build and operate systems that run machine learning models in
production, and I write about what I find while doing it.

Most of that work sits at the boundary between a model and everything around it:
the serving layer, the memory budget, the evaluation harness, the part of the
pipeline that quietly decides what is possible.

What I work on#

Day to day, that means inference performance, agent architectures, and the
infrastructure underneath both. I spend a lot of time with profilers, a fair
amount with papers, and more than I would like with GPU memory allocators.

Why I write#

Two reasons.

The first is selfish: writing an explanation is the fastest way I know to find
out whether I actually understand something. Most of these pieces started as a
note to myself that turned out to be wrong somewhere in the middle.

The second is that the field moves quickly and explains itself badly. There is
an enormous amount written about what AI systems do and comparatively little
about how — the mechanics, the trade-offs, the numbers. That gap is where I
try to write.

How this site works#

Articles are written in Markdown and stored as Markdown — no proprietary format,
no lock-in. Every claim that comes from somewhere else has a reference at the
bottom of the piece. Measurements include the hardware they were taken on.

There are no trackers, no advertising, and no third-party scripts. Analytics
here means a daily-rotating hash that counts a view once and cannot be linked
back to a person. Likes work the same way: a random id your browser keeps, which
the server only ever sees as a salted hash.

Find me#

The fastest way to reach me is LinkedIn — I read everything that comes in, and
corrections are genuinely welcome, particularly on the benchmark numbers.

• LinkedIn

If something here was useful, or wrong, I would like to know which.