Hi! 👋 I'm Arpan.

FB Live Video Experience, managed by Mr. Gabriel Ochoa

• Created infinite scroll comments on WWW using React, PHP, and Javascript and live polls video overviews on iOS and Android generating a sig-stat increase in FB Watch time based on A/B testing over 200+ million users.

• Simplified E2E testing framework and documentation for internal languages used by 50+ teams. Received greatly exceeds expectation (GE) performance rating.

FB Shops Ranking Team, managed by Mr. Artem Zinchenko

• Built a product retrieval transformer model personalized to users based on their history and viewing habits. This model beats the current production retrieval model by 35%.

• Optimized the existing collection pipelines to operate 600x faster using Hack, PHP, and Python. Created a new API to help engineers query PyTorch models which is currently used as an internal tool.

IMPACT Lab, advised by Prof. Wen-mei Hwu and Dr. Vikram Sharma Mailthody

• Generated the largest publicly available graph dataset - Illinois Graph Benchmark (IGB) with 600M nodes and 6B edges collaborating with researchers from Amazon AWS, Deep Graph Library (DGL), NVIDIA, and IBM for supervised node classification and efficient system designing.

• Combined Microsoft Academic Graph (MAG) and Semantic Scholar databases to annotate 162× more data for supervised learning tasks to test emerging Graph Neural Network (GNN) models at scale.

Undergraduate Research
January 2022 - Present

Forward Data Lab, advised by Prof. Kevin Chen-Chuan Chang

• Created a 2-stage attention-based seq2seq model to generate subtopics for a given title that performs better than SOTA models for fine-grain attribute tagging. Filtered out noisy web retrieved data using text-rank and trained a few-shot classifier model to classify them into the dynamically generated sub-topics.

IGL Fellow @
Illinois Geometry Lab
January 2022 - Present

Hildebrand Research Group, advised by Prof. AJ Hildebrand

• Performed large-scale statistical analysis of the first 30 billion continued fraction (CF) digits of π using chi-square tests, p−tests, and Kolmogorov–Smirnov tests to provide compelling evidence that these digits behave like those of a random real number.

• Implemented massive multi-dimensional space walks to compare random and π CF digits and used extreme-digit, single-digit, and z − scores for further evidence.

SPIN Intern @
National Center for Supercomputing Applications (NCSA)
September 2021 - May 2022

HathiTrust Digital Library, advised by Prof. J. Stephen Downie and Dr. Glen Layne-Worthey

• Applied transfer learning on a Mask-RCNN based model detectron2 using the Pubmed dataset to detect 5 non-text classes in documents. Coupled with a large image classification model to detect and classify over 1000 classes of non-text objects and images on scanned documents with over 97% accuracy.

• Used openCV for preprocessing pipeline to process over 16M volumes (5B pages) in the HathiTrust Digital Library to improve run-time by 3× on a V 100 GPU.

IGL Fellow @
Illinois Geometry Lab
August 2022 - Present

Social Computing Lab, advised by Prof. Jana Diesner

• Using Twitter API to collect 10M tweets for detecting and prioritizing needs during crisis events using NLP (like the Russia-Ukraine conflict) in order to (1) extract a list of needed resources, (2) how they are fulfilled.

Undergraduate Research
October 2020 - May 2021

Koyejo Lab, advised by Prof. Sanmi Koyejo

• Worked on a cross-department project to predict phenotype combinations in maize/sorghum crop genes to maxi- mize heritability using reinforcement learning (RL). Wrote classic local and global search algorithms like Particle Swarm Optimization (PSO) and Simulated Annealing (SA) to set baselines.

• Developed a Multi-Layer Perceptron (MLP) to serve as a mapping function between wavelength and experimental ground-truth data which improved the RL search over multi-dimensional wavelength space.

Undergraduate Research @
OSF Healthcare-UIUC Jump ARCHES
June 2020 - May 2021

Health Care Engineering Systems Center (HCESC), advised by Dr. Inki Kim

• Developed a novel operation training procedure in virtual reality (VR) using Unity with anatomically accurate physics scripts, capable of real-time rendering optimized to run without a GPU.

• Implementing a reinforcement learning (RL) model to automate operation procedure and create a predictive model to assign probability of success using 6-dimensional c-space A*, RRT, and RRT* search algorithms.

Undergraduate Research
September 2020 - May 2021

Caesar Lab, advised by Prof. Matthew Caesar

• Experimenting combinations of different neural net architectures for object detection (ResNets, MobileNets) and object tracking using deepSORT with an SSD trained on NCSA’s HAL cluster.

• Implemented the Hungarian Algorithm and Kalman Filter to track and predict object position during large periods of obstruction.

Undergraduate Research @
EarthSense
October 2019 - March 2020

FRESH Lab, advised by Prof. Girish Chowdhary

• Soldered and worked on the power source and circuit design for cameras and sensors of an agricultural bot.

• Wrote Python scripts for autonomous path planning using OpenCV. Labeled over 5000 pictures and helped train a Mask R-CNN model to detect the path and its surrounding with less than 5% error.