hi! I'm Shiv, I'm a programmer based in London. I'm interested in deep learning, specifically mechanistic interpretability, and I'm currently exploring projects in this space. when I'm not coding or doing maths, I like hiking and making music.
if you'd like to get in touch, feel free to email me at shivbhatia10 at gmail dot com.
# these are some of the things I've worked on (the date is when I finished it):
[jan 2025] - organised a hackathon focused on interpretability in AI as part of SoTA, a non-profit I co-founded with the aim of promoting techno-optimism in the UK.
[sept 2024] - came back for a second internship at Palantir, where I developed a version control system for AIP products.
[june 2024] - graduated with a first class MEng from Imperial College London where I studied maths and CS, mostly geometry, quantum physics, and deep learning.
[oct 2023] - spent a summer as a software engineering intern at Palantir, where I worked on automations for ontology entities.
[oct 2023] - led computer vision on the only UK team to compete in the world finals of the European Rover Challenge 2023. we built a Mars rover prototype with a 360 degree camera on its head that was capable of simultaneous localisation and mapping on an uneven terrain. ranked 1st on one of the major challenges and 11th out of 25 global teams in the final. media coverage
[june 2023] - built RCoin, a real time auditable stablecoin pegged to the South African Rand, along with my flatmates. we built a complete app and working payment platform which was designed to replace expensive and monopolistic payment solutions available in Eswatini. got to consult directly with the finance minister on the design of the platform. we were the headline presentation at the Brevan Howard crypto showcase 2023.
[oct 2022] - finished an internship at Microsoft, where I worked on adding video conferencing to a unified communications app.
# here is a graveyard of many half-finished projects and some completed ones:
[wrist] - a fast, shell-agnostic terminal history application for MacOS written in Rust.
[torch_sparse_autoencoder] - a minimal Pytorch library for handling sparse autoencoders for language/ing a simple API.vision models. designed to work for a wide range of model architectures while still providing a simple API.
[torch_sparse_autoencoder] - a minimal Pytorch library for handling sparse autoencoders for language/vision models. designed to work for a wide range of model architectures while still providing a simple API.
[deto_x] - a Chrome extension that lets you filter out tweets with a custom prompt and LLMs.
[IR spectra CNN classifier] - scraped a bunch of IR spectra data for organic compounds and trained a convolutional neural network to detect functional groups. the final model achieved an F1 score better than an average chemistry undergrad at Imperial.
[random cellular sheaf generator] - often in computational algebraic topology, it's useful to be able to randomly generate simplicial complexes of arbitrary dimension in a usable format for experiments. working with researcher at Imperial I wrote a small library for this purpose, specifically tailored to persistent homology experiments for topological data analysis.
[csvtui] - a small Rust app that lets you view and edit csv files from the terminal with Vim-like keybindings.
# TODO: add the rest