Skip to content
View aorko01's full-sized avatar
20 followers · 19 following

Highlights

  • Pro

Block or report aorko01

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
aorko01/README.md 
███████╗██╗   ██╗██╗     ███████╗██╗██╗  ██╗███████╗██████╗
╚══███╔╝██║   ██║██║     ██╔════╝██║██║ ██╔╝██╔════╝██╔══██╗
  ███╔╝ ██║   ██║██║     █████╗  ██║█████╔╝ █████╗  ██████╔╝
 ███╔╝  ██║   ██║██║     ██╔══╝  ██║██╔═██╗ ██╔══╝  ██╔══██╗
███████╗╚██████╔╝███████╗██║     ██║██║  ██╗███████╗██║  ██║
╚══════╝ ╚═════╝ ╚══════╝╚═╝     ╚═╝╚═╝  ╚═╝╚══════╝╚═╝  ╚═╝

ML Infrastructure · GPU Programming · Systems Engineering

LeetCode Codeforces LeetGPU

LinkedIn Medium dev.to

> whoami

I build things close to the metal — GPU kernels, distributed training infrastructure, and systems software. I care about what happens at the layer most engineers treat as a black box: how memory moves, how schedulers make decisions, how kernels actually run on hardware.

Currently going deep on ML systems, GPU programming, and low-level systems (OS internals, networking, memory). Previously built full-stack apps; now I live in the backend.

> ls ./projects

🔧 Distributed ML Task Scheduler

In Progress

A cluster-level job scheduler for heterogeneous GPU environments.

  • Users submit a training script + dataset via a simple web UI — the system handles the rest
  • VRAM & runtime estimation across different GPU models before scheduling
  • Bin-packing: multiple jobs share a single GPU when headroom allows, maximizing utilization
  • Checkpoint-based fault recovery: if a node dies, the job resumes from last checkpoint
  • Job feedback: users get stdout/stderr on completion or failure
  • Tech: Python · Ray · Docker · Kubernetes · Prometheus

⚡ Flash Attention in Triton

In Progress

Implementing Flash Attention from scratch using Triton GPU kernels, then training GPT-2 on top to validate correctness and benchmark real-world throughput.

  • Tiled SRAM-based attention to avoid HBM round-trips
  • Benchmarking against standard PyTorch attention
  • End-to-end GPT-2 training to verify gradient correctness
  • Tech: Triton · CUDA · PyTorch · Python

> cat skills.txt

Languages

Python      ████████████ proficient
C++         ████████████ proficient
Rust        █████░░░░░░░ learning
Go          █████░░░░░░░ learning

ML & GPU

PyTorch         ✓ proficient
PyTorch DDP     ✓ proficient
CUDA            ✓ proficient
Triton          ✓ proficient

Infra

Docker          ✓
Kubernetes      ✓
Ray             ✓
Prometheus      ✓

Systems

OS Internals    → active
Networking      → active
Memory Mgmt     → active

Previously

React / Django
PostgreSQL / MongoDB
Full-stack web

DSA

LeetCode grind
Codeforces
GPU puzzles @ LeetGPU

> github --stats

Pinned Loading

  1. Distributed_ML_Training_scheduler Distributed_ML_Training_scheduler Public

    Centralized scheduler that runs user-submitted ML training jobs on heterogeneous GPUs, optimizes GPU utilization, handles failures with checkpointing, and provides real-time feedback via a web inte…

    Python

  2. RemoteWake RemoteWake Public

    Remotely power on and shut down an Ubuntu desktop over the internet using an ESP32, Wake-on-LAN, and Cloudflare Tunnel — zero recurring cost.

    C++

  3. MiniTorch MiniTorch Public

    A mini torch like library written in cpp from scratch and binded to python

    C++

  4. Flash_GPT_2 Flash_GPT_2 Public

    Custom GPT-2 with memory-efficient FlashAttention implemented in Triton for faster training in PyTorch.

    Python

  5. Thread_Pool Thread_Pool Public

    C++

  6. Chrome_extension Chrome_extension Public

    JavaScript