top of page

ThunderKittens

An embedded CUDA DSL + fast and clean kernels

Screenshot 2024-05-12 at 4.24.07 PM.png

leveraging multi-agent reinforcement learning (marl) to enhance e-amod fleet control

we built on an exisiting tri-level framework that utilizes graph-rl to solve the e-amod fleet control problem. we test how well a marl approach performs given a situation where vehicles in an e-amod fleet need to be continuously recharged and spatially rebalanced.

Screenshot 2024-01-12 at 2.27.58 PM.png
Bitcoin

marabu

we built a blockchain protocol from scratch

tuning video segmentation for creating masks of pedestrians

we used fair’s recent segment anything model (sam) to create masks of pedestrians from driving footage. fine-tuning sam for pedestrians can lead to easier applications of machine learning models for pedestrian agent prediction that would otherwise be too computationally expensive to train on video data

Screenshot 2024-01-12 at 2.30.44 PM.png

stanford treehacks 2023

project: condensis

my team and i built condensis in 48 hours - our web-app utilizes openAI's api to generate condensed lecture notes from videos of college lectures - the multimodal language model powering it can synthesize high-quality, easy-to-understand notes from any lecture video

Screenshot 2023-02-20 at 4.04.27 PM.png

musestep

we built a piano with stepper motors

Screenshot 2024-01-12 at 2.37.43 PM.png
bottom of page