Jivitesh Jain
Tagline:Graduate Researcher at Carnegie Mellon. All things NLP, LLMs, and Responsible AI.
Pittsburgh, PA, USA
Hello
👋 I'm a research masters student in Language Technologies (MLT) at Carnegie Mellon University in the steel city of Pittsburgh. My current research with Prof. Mona Diab revolves around Responsible AI, with projects ranging from using interpretability techniques to predict the chances that an LLM is hallucinating (which, in turn, gives us confidence while using LLMs in factuality-critical applications) to the use of LLMs for regulatory compliance of software. I'm also interested in training and fine-tuning LLMs, building end-to-end workflows such as RAG, and making LLMs run efficiently on smaller devices.
Previously, I spent two amazing years in London as a software engineer at Palantir. There, I built high-scale, high-availability backend systems that enabled data and ML workflows in Palantir Foundry and AIP -- Palantir's super capable platforms for doing cool (and useful) things with your organization's data using AI. This, this, this, and this are examples of projects that I led, worked on, or contributed to.
Before that, I graduated with honors [, two gold medals, and many lifelong friends] from IIIT Hyderabad with a bachelors degree in Computer Science. While there, I worked on problems in computational social science with Prof. Ponnurangam Kumaraguru and Prof. Joyojeet Pal at the Precog Research Group.
I also explored the field of 3D computer vision at Brown University’s Interactive 3D Vision and Learning Lab (IVL) with Prof. Srinath Sridhar. Our work dealt with the category-level pose canonicalization of full as well as partial objects using tensor-field networks. During that time, I also interned at Google and participated in Major League Hacking’s Fellowship. Look around here to find projects I’ve worked on or my involvements in hackathons and programming competitions.
After slamming down my laptop's lid, I like to spend time creating art, discovering the best hot chocolate places in town, and sipping hot chocolate while creating art.
If you find me a good fit for your team, want help, or just want to say hello – please freely reach out via the links below!
Publications
Condor: Self-Supervised Canonicalization of 3D Pose for Partial Shapes
Conference PaperPublisher:IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)Date:2022Authors:Rahul SajnaniAdrien PoulenardJivitesh JainRadhika DuaLeonidas J GuibasSrinath SridharDescription:Progress in 3D object understanding has relied on manually “canonicalized” shape datasets that contain instances with consistent position and orientation (3D pose). This has made it hard to generalize these methods to in-the-wild shapes, e.g., from internet model collections or depth sensors. ConDor is a self-supervised method that learns to Canonicalize the 3D orientation and position for full and partial 3D point clouds. We build on top of Tensor Field Networks (TFNs), a class of permutation- and rotation-equivariant, and translation-invariant 3D networks. During inference, our method takes an unseen full or partial 3D point cloud at an arbitrary pose and outputs an equivariant canonical pose. During training, this network uses self-supervision losses to learn the canonical pose from an un-canonicalized collection of full and partial 3D point clouds. ConDor can also learn to consistently co-segment object parts without any supervision. Extensive quantitative results on four new metrics show that our approach out-performs existing methods while enabling new applications such as operation on depth images and annotation transfer.
Projects
Implementation and Analysis of state-of-the-art Retrieval-Augmented Generation Techniques for Question Answering
date: 2024Organization:Carnegie Mellon University
Description:This project explores all aspects of Retrieval-Augmented Generation (RAG), including collecting data for the knowledge corpus, generating synthetic Q/A pairs on that data for parameter-efficient quantized model fine-tuning using LoRA, annotating a gold test set, using SOTA embedding and indexing methods, retrieved document reranking and summarization, and query rewording using Hypothetical Document Embeddings (HyDE). We perform comprehensive ablation studies to analyze the effect of each component of the pipeline in detail, and release our dataset and code publicly.
Mini Llama
date: 2024Organization:Carnegie Mellon University
Description:A miniaturized implementation of the Llama 3.1 Language Model in PyTorch, with Q-LoRA parameter-efficient fine-tuning, trained on the TinyStories dataset.
SimpleRA Database Management System
date: 2021Organization:International Institute of Information Technology, Hyderabad
Description:A DBMS implemented from scratch in C++, with query parsing and planning, linear hash and B+ tree indexing, 2-way merge sort, caching and memory page management.
Multilingual Sparse Indexing and Search for Wikipedia
date: 2021Organization:International Institute of Information Technology, Hyderabad
Description:An efficient search and indexing engine for English and Hindi Wikipedia with blocked-sort indexing and distributed search to reduce index size by 75% and search 80GB of XML in <1 second.
Image Segmentation and Background Removal using GrabCut
date: 2021Organization:International Institute of Information Technology, Hyderabad
Description:This project implements major portions of the GrabCut Algorithm, a markov random field based image segmentation algorithm, a variant of which you might have seen available in Microsoft Office Products as the background removal tool. The algorithm is iterative and interactive, and is known for the minimal and simple user interaction required, as well as its ability to fix the results using further iterations of user input. The algorithm is based on multiple iterations of GraphCut, and uses each iteration to improve its model of the foreground and background color distributions, which are modeled as gaussian mixtures.
Citation Intent Classification
date: 2021Organization:NAACL
Description:An NLP system to classify citation intent in scientific literature using fine-tuned transformer models like RoBERTa, for the 3C SDP shared task at NAACL 2021, implemented using PyTorch.
Kishmish - A shell for Linux
date: 2019Organization:International Institute of Information Technology, Hyderabad
Description:A bash-like shell for Linux systems, implemented from scratch in C, with support for process management, piping and redirection and built-in as well as system commands.
Improvements to the XV-6 Operating System
date: 2019Organization:International Institute of Information Technology, Hyderabad
Description:Improvements to MIT’s xv-6 operating system with additional system calls, and implementations of FCFS, Priority-Based and Multi-Level Feedback Queue process scheduling, written in C.
Resume
DownloadContact
Address
5412 Gates and Hillman Centers
Carnegie Mellon University
4902 Forbes Ave
Pittsburgh 15213, PA