Gargi Gupta

hi, i'm gargi.

@vjti, mumbai

SRA Webring

someone who loves understanding how intelligent systems think, learn and move.

i'm 18, and a part of the SY core @SRA VJTI, which allows me to work with people who think fast, argue well and deliver most ideas into execution.

I spend most of my time trying to understand -

how an algorithm chooses,
how a robot moves,
how a model learns from noise.

currently exploring

reinforcement learning machine learning computer vision competitive programming finance

things that fuel me

📚 books 🎵 music 🎨 art

projects.

a glimpse into my exploratory ventures.

QuadMove

completed

RL curriculum ending in PPO on a Unitree Go2 in MuJoCo. spent most of the time on reward design. a policy can pass every metric and still look completely wrong.

github

AmpMove

in progress

AMP on iRonCub in Isaac Lab. instead of a hand-crafted reward, a discriminator scores how closely the policy's motion matches CMU mocap clips and that score becomes the reward signal.

github paper

TinyDiff

in progress

scalar autograd engine from scratch. builds the graph forward, walks it back. extending to tensors.

github micrograd lecture

ViTs on CMS Detector Data

completed

benchmarked L²ViT and PolaFormer on CMS detector images. L²ViT matches softmax at O(n). PolaFormer's edge only shows after pretraining, ReLU kernel is the bottleneck on sparse regression.

github L²ViT paper PolaFormer paper

implement_Eureka

in progress

Eureka generates reward functions from env code using GPT-4. we add visual goal specification via CLIP and swap GPT-4 for a local LLM, so you can show the agent what to do rather than only describing it in code.

github paper