Hi, I’m Abhik Sarkar.
I’m the Director of Machine Learning at Cloudastructure Inc, where I scaled our computer vision platform from 100K to 12 million videos a day and look after the on-prem GPU cluster that runs it.
If you came here from an article or a paper review: hi, I wrote it. This site is where I explain what I learn along the way, from GPU optimization and transformers to video codecs and the papers behind them, usually with something you can click and poke at.
Follow along
Start here
Pick the format you like. Each one starts with a piece worth reading first.
Articles
Long-form write-ups from production work: GPUs, video codecs and Python internals.
How TensorRT WorksConcepts
Short interactive explainers on ML, GPUs and systems, grouped by topic.
Flash Attention: IO-Aware Exact AttentionPapers
Reviews of machine learning papers, with demos you can run in the browser.
ViT: An Image is Worth 16x16 WordsMy story
What I do now
At Cloudastructure I use deep learning and computer vision to improve safety and security. Day to day that means Python, C++, TensorRT, PyTorch, CUDA, Docker and React.js, plus MLOps tooling and distributed systems for large-scale video processing. I also hire and lead the team, which spans 11 countries, and I talk about this work at conferences.
Spoken at
All talks- EuroPython
- PyCon Italia
- PyCon India
- PyCon Japan
- BangPypers
- DevFest
Why GPUs
I’m an accidental performance engineer.
It started as a way to save the company money. Then I learned that decoders like NVDEC are separate silicon from the CUDA cores, so decode, preprocessing and inference can run on different parts of the GPU at the same time. Pipelining them gave us a 3x speedup and saved thousands in infrastructure costs at scale. I’ve been hooked since: CUDA kernels, memory access patterns and TensorRT graph optimization, all the way down to streaming multiprocessor occupancy.
Beyond work
I mentor, run workshops and write here to share what I learn. Away from the keyboard you’ll find me at quizzes or trying out whatever new tool just shipped.
Things I’m passionate about
- Machine Learning
- GPU Performance Engineering
- Software Engineering
- Python
- Open Source
- Quiz
- Data Engineering
How I got here
I have 8+ years of experience as a machine learning engineer and full-stack developer. My training is in computer science: a Bachelor of Technology in Computer Science and Engineering from the National Institute of Technology, Raipur, India, followed by graduate coursework in Introduction to High-Performance Computing at the Indian Institute of Science, Bangalore, and CS224W: Machine Learning with Graphs through the Stanford Center for Professional Development. The full list is in my résumé and credentials, and my expertise areas map the topics I cover.
Articles I Endorse
A selection of articles that have influenced my thinking and approach to technology.
- Making Deep Learning go Brrrr From First Principles — Horace He
- Transformers from Scratch — Brandon Rohrer
Say hello
Email is the best way to reach me, whatever the reason.
Email me