Zafar Takhirov
Founder AI Researcher at Stealth Startup
Mountain View, California, United States
Summary
🤩
Rockstar🎓
Top SchoolZafar Takhirov is a founder and AI researcher with 13 years of experience building efficient, privacy-preserving generative AI systems from research prototypes to production-focused implementations. He has led AI-privacy work at TikTok and implemented quantization and pruning kernels at Meta, bringing low-level CUDA/CPU expertise together with model compression and sparsity research. A longtime contributor to PyTorch, he has improved sparsity and quantization tooling—including a cubic sparsity scheduler and block-pruning fixes—bridging open-source frameworks and production needs. His background spans academia (PhD work on energy-efficient, fault-tolerant ML systems) to hands-on mixed-signal and SoC engineering, giving him a rare blend of hardware-aware ML and software engineering. Based in Mountain View, he now runs a stealth startup focused on making genAI both private and efficient, and he enjoys translating theoretical ideas into pragmatic systems for industries that haven't yet caught up with the AI hype.
13 years of coding experience
14 years of employment as a software developer
Bachelor of Science / Scpecialist, Linguistics, Bachelor of Science / Scpecialist, Linguistics at Russian-Tajik Slavic UniversityDoctor of Philosophy (PhD), Computer Engineering, Doctor of Philosophy (PhD), Computer Engineering at Boston University
Tajik, Turkish, German, Russian, English
Github Skills (10)
neural-network10
quantization10
sparse10
pytorch10
machine-learning10
deep-learning10
python10
gpu7
autograd7
tensor7
Programming languages (14)
Github contributions (5)
pytorch/pytorch
Apr 2019 - Oct 2022
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Role in this project:
ML Engineer
ML EngineerContributions:273 reviews, 799 commits, 454 PRs in 3 years 7 months
Contributions summary:Zafar primarily contributes to the PyTorch framework, focusing on enhancements related to sparsity and quantization techniques for deep learning. They refactor and move code related to sparsity, including experimental features and modules. The user introduces a mechanism for lazy deprecation imports within the nn module and improves the handling of modules during migration. Additionally, they implement a cubic sparsity level scheduler and address issues related to quantization engines and shapes for block pruning.
pythongpu-accelerationdeep-learninggpunumpy
jdansey/EC311-Project
Dec 2015 - Jun 2021
snake game
Contributions:2 PRs, 1 issue in 5 years 6 months
snakesnake-gamejavascriptgame
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.