colorscheme snowclave

Vadodara

~/hello

himanshu
sharma

i build the things underneath my tools to see how they work: an autograd engine in C++ and CUDA, a Mandelbrot renderer, an RL benchmark. the background is that renderer, running live.

1234 switch workspaces

fastfetch
himanshu@vadodara
Now
Web dev, Indian Rocketry Assoc.
Before
ML intern, ThinkandForm (NZ)
Host
GSFC University, B.Tech CSE
Learning
Reinforcement learning
Langs
C++, CUDA, Python, Rust, Lua
OS
Arch Linux, Hyprland
colorschemefrom my dotfiles

    ~/src $ ls

    work + school

    work

    1. summer 2026

      ThinkandForm ML intern

      Trained U-Net models to segment hair and scalp, in pipelines anyone can rerun: Docker, DVC, Weights & Biases, RunPod.

    2. 2026 to now

      Indian Rocketry Association web developer

      I'm building a 3D investor site: a scroll-driven story with live rocket launch simulations and trajectory plots, in React, Three.js and GSAP.

    school

    1. 2025 to 2029

      GSFC University B.Tech CSE

      In Vadodara. Coursework covers linear algebra, probability, OS and architecture. On the side I work through Stanford's CS231n and CS224n, and DeepMind's RL lectures.

    older stuff lives on linkedin.

    source
    ensembleqHumanoid-v5, CrossQ

    PPO, SAC, TD3, FastTD3 and CrossQ, benchmarked on MuJoCo and Fetch with one GPU and checkpoints every 50k steps. The video is CrossQ on Humanoid-v5, ~11,100 return at 1.5M steps. SAC reached ~5,400 at 1M.

    rl-botCoppeliaSim, 2x speed

    A line-following robot trained with Q-learning in CoppeliaSim for e-Yantra at IIT Bombay. I then raced it against a PID controller I tuned by hand.

    source
    enerscripttype here

    My first language, a BASIC-flavoured toy. I wrote the lexer, parser and evaluator by hand. Put TOKENS or AST in front of a line to see inside.

    source
    fractalsdrag, scroll to zoom

    A Mandelbrot renderer in C++ on OpenGL 3.3. Ported to WebGL with double-float maths (emulated double precision), so it stays sharp at deep zoom.

    source
    energrad

    An autograd engine in C++ and CUDA. It records the graph as it computes, orders it with Kahn's algorithm, and runs the backward pass on custom CUDA kernels.

    C++ CUDA

    source
    cuda-sgemm

    Single-precision matrix multiply in hand-written CUDA kernels. It was the warm-up for energrad's GPU side.

    CUDA C++

    source
    clumsy-cluster

    A multi-agent RL environment that simulates a compute cluster, built on Alibaba's cluster traces with PettingZoo.

    Python PettingZoo

    source

    more on github

    ~/.config $ ls rices

    on the side i rice my Arch Linux + Hyprland desktop. pick a rice and the whole page recolours to match it.

    waypaper1 / 1

      Contact

      type to filter, tab switches modes, enter to open

      the background is my fractal renderer. the colours come from my dotfiles.