Vedant Kulkarni

About Me

I am an undergraduate computer science student at IIIT Hyderabad. My current areas of interest are mechanistic interpretability and diffusion models.

I worked on AgentTorch with MIT Media Lab’s Large Population Models team in 2024, and on function calling in the Gemma 3 family of models with Google DeepMind in 2025, as a Google Summer of Code project. In 2026, I am working with Davide Baldelli on Building a Robust Activation Monitor, a SPAR fellowship project on activation monitors that stay reliable across models, concepts, and contexts.

I have been writing open source code since 2018, and have contributed across 400+ issues and pull requests in 30+ repositories on GitHub. I co-maintain the popular node.js libraries express-rate-limit and express-slow-down, used to shield servers from DDOS attacks, with their original author Nathan Friedly. I am also a contributor to serve, Refined Github, FPM, Beckn, Sunbird Registries and Credentials. I co-created the Mess Portal and associated systems at IIIT Hyderabad with Aarnav Pai, and continue to maintain them.

I participated in the national ONDC Grand Hackathon in 2022 and my team won a special mention award. I was invited to give a talk titled ‘A 12 Grader’s Journey to Opensource Contribution via Sunbird and Beckn’ at the Sunbird DPG Tech Fusion Conference, Pune, in 2024. I also gave a talk on my work on AgentTorch at the FOSS United Meetup, Hyderabad, in 2024.

Highlights

A few of the projects I have created, maintain, or contributed to. The full list has everything.

AI and Machine Learning

offline-function-calling

created, GSoC 2025 with Google DeepMind

My Google Summer of Code 2025 project with Google DeepMind, exploring, extending, and documenting the function calling capabilities of the Gemma model family. Over 12 weeks, the project grew from simple experiments to a benchmarking suite and leaderboard covering 25 Gemma variants and quantizations, a function calling SDK, a command-line interface not unlike the Gemini CLI, and a set of tutorials for developers working with offline models.

agent-torch

contributed, part of the Large Population Models project by MIT Media Lab

Created a custom model using the framework to simulate movement of predator and prey on a map of Central Park, NY, and added a feature that allows for easier declaration of substeps using decorators. Worked on improving the Python API, creating interactive simulations for use cases such as solar panel adoption and zoning policy, and integrating it with Beckn.

Libraries and Web Infrastructure

express-rate-limit

co-maintainer3.3k stars220M dl/month

Rewrote the library in Typescript, created the modern store interface and continue to maintain it alongside Nathan Friedly, the original author. Help in maintaining 6 out of the 8 stores written for the library, and helped rewrite the default memory store to make it more memory-efficient.

express-slow-down

co-maintainer303 stars610k dl/month

Helped with the rewrite in Typescript, and achieving feature parity with express-rate-limit. I continue to maintain it alongside Nathan Friedly, the original author.

serve

contributed9.9k stars14M dl/month

Refactored the library, added tests, and transitioned to ESM + Typescript.

Developer Tools and Systems

refined-github

contributed32.2k stars80k+ users

Contributed to 8 popular features, including useful-not-found-page, restore-file, quick-comment-edit, linkify-code, and forked-to.

fpm

contributed11.5k stars

Restructured the documentation, rewrote parts of the installation page, added a getting started page, tutorials, and examples, as well as a CLI flags reference page.

Digital Public Goods

sunbird-rc

contributed

Documented the process of creating and working with registries, and created the registry-cli package to spin up a registry quickly and easily. Submitted various RFCs to improve the project.

beckn

contributed

Helped create a site for developer documentation using Gatsby and Astro, and dockerized the reference Beckn app implementation.

verifiable-presentations

created

The specification and minimal reference implementation for a plugin-based service that allows issuers to render verifiable presentations, based on the W3C verifiable credential specification from templates, and store it in a queryable database for holders to list/retrieve.

Applications

mess-portal

co-creator and maintainer

Students at IIIT Hyderabad use it to register for the mess and check their bills. It automates a lot of the administration’s paperwork, such as billing. I co-created it with Aarnav Pai, and continue to maintain it and the associated systems.

Areas of Interest

These are the things I am working on at the moment.

Mechanistic interpretability, which means reverse engineering the features and circuits that neural networks learn, to understand why models behave the way they do. I am most interested in its applications to AI safety and alignment. A model whose internals you can read is one you can check for deception or misalignment before it acts. My SPAR project on activation monitors is a step in that direction.
Diffusion models, how they learn and represent concepts, and their practical applications beyond image generation, such as in diffusion-based language models.

These are areas I have explored before and really loved.

Multimodal function calling with LLM agents, and benchmarking the function calling capabilities of different model sizes and quantizations. This grew into my GSoC project.
Agent based modelling with LLMs, which I worked on as part of AgentTorch.
Custom Android ROMs. I customised and compiled LineageOS for my own phone.
Low level languages, to better understand how the computer works. I followed tutorials to write a kernel in assembly and Rust, including Philipp Oppermann’s.