Aayush Garg
  • Writing
  • Work
  • Tools
  • About

Work

Projects, publications and live tools. Most of the code is on GitHub.

Live

  • India Weather

    Live temperature, humidity, and air quality across India's major metropolitan cities.

  • Browser tools

    Image, viewer and data utilities that run entirely in your browser.

No matching items

Open source GitHub

  • building-from-scratch

    My ongoing repo where I build language models from first principles. So far:

    • Character-level GPT
    • GPT-2 124M reimplementation with RoPE
    • BPE tokenizer from scratch (with a Rust encoder for 3.7x speedup)
    • LLM inference with KV cache and speculative decoding
    • Supervised fine-tuning for reasoning and instruction following
    • Expert iteration for self-improvement on math
    • Group Relative Policy Optimization (GRPO)

    Currently working on a FlashAttention2 implementation from first principles.

  • cv-skills

    A Claude Code plugin with skills for classical image processing: format conversion, SVG-to-raster, resize/crop/rotate, color adjustments, blur/sharpen/denoise, edge detection and segmentation, and image compositing with background removal.

  • ComfyUI-Svg2Raster

    I wrote this ComfyUI custom node package for loading SVG files and rasterizing them into images with configurable width, scale, background color, and border. Uses CairoSVG under the hood.

  • tutorials

    Tutorials I wrote for myself and shared along the way.

  • spatial-alias-removal

    I used deep learning super-resolution (SRCNN, EDSR, VDSR) to upscale seismic data by 2x in the spatial direction and remove spatial aliasing. This work was published as a paper back in 2019, built during my time at TU Delft and University of Amsterdam.

  • research-papers

    My reading notes and summaries for deep learning papers organized by category. Covers diffusion models (DDPM, DDIM, ControlNet, SDXL, and more), Transformers (Attention Is All You Need), and GANs.

  • pytorch-pl-hydra-templates

    I put together four progressively more sophisticated PyTorch training templates using CIFAR-10: vanilla single-GPU, multi-GPU with DistributedDataParallel, PyTorch Lightning with TensorBoard, and PyTorch Lightning + Hydra for rapid experimentation with swappable models, optimizers, and configs.

  • model-parallelism

    I explored how to pipeline models with skip connections (ResNets, UNets) across multiple GPUs using torchgpipe.

  • finance_dashboard_example

    A sample Plotly Dash interactive dashboard I built for tracking stock price movement with simple moving average and candlestick pattern charts.

No matching items

Gists

  • Step-by-Step Guide to setup your own personal GPU server

    A guide to setting up remote SSH access to a personal GPU workstation running Ubuntu, covering SSH keys, firewall setup, and ngrok tunneling.

  • Managing multiple CUDA versions using environment modules

    My notes on installing multiple CUDA versions side by side on Ubuntu and switching between them using `environment-modules`.

No matching items

Publications All publications

Journal papers and conference talks from my PhD and research years.

© 2026 Aayush Garg

 
  • GitHub

  • LinkedIn

  • X

  • RSS