Sai Varun Reddy, Mullangi

Graduate Research Assistant

Arizona State University

Biography

I’m a CS master’s student with a passion for turning data into actionable insights and crafting robust software solutions. With three years of diverse work experience, I’ve honed my skills as a software developer and data scientist.

My expertise spans multiple programming languages including Python, Java, C/C++, JavaScript, and SQL. I’m well-versed in utilizing frameworks and tools such as PyTorch, TensorFlow, REST, Flask, Redis, MongoDB, GNU, GDB, Git, Kubernetes, and Docker to build cutting-edge applications.

Currently pursuing my master’s, I’m eagerly seeking full-time opportunities starting in 2024 to apply my knowledge and skills in a dynamic work environment.

Interests

Natural Language Processing
Application Development
Large Language Models

Education

MSc in Computer Science, 2024
Arizona State University
MSc in Mathematics, 2019
Birla Institute of Technology and Science (BITS), Pilani
BE in Mechanical Engineering, 2019
Birla Institute of Technology and Science (BITS), Pilani

Skills

Technical

Python

Data Science

SQL

Hobbies

Football

Hiking

Coding

Experience

Data Scientist

ZS Associates

January 2021 – Present Bangalore

Projects include:

Part of a team that designed a positive-unlabeled machine learning algorithm to identify ovarian cancer patients with missing insurance claims for surgery, biomarkers, and PARP (poly ADP-ribose polymerase inhibitors) which resulted in identification of $40 million unaddressed opportunities for the client.
Developed a Seq2Seq LSTM Encoder-Decoder model to identify anomalous ovarian cancer patient journeys and used cross-entropy loss to evaluate the anomalous patient cluster.

Machine Learning Engineer

Quantiphi Analytics

January 2019 – June 2019 Mumbai

Projects include:

Implemented containerization of an existing solution and deployed it on a Google Kubernetes Orchestration Cluster, while also building a highly scalable video processing framework using Pub/Sub, resulting in a 158% improvement in the client’s product efficiency
Part of a team that designed and implemented a binary CNN based AI model for top US medical firm to classify images of a SARS-CoV-2 Antigen test cassette for the COVID-19 virus, achieving exceptional sensitivity and specificity of 0.99 and 0.98 respectively
Developed and implemented a data retrieval model that fetches information from two separate unstructured databases for a given user query. Utilized Universal Sentence Encoder to create embeddings, weighting column names and values, and integrated DistilBERT to find the exact answer from the top 3 scored columns based on their Cosine Similarity

Projects

xv6 Memory Management Mastery - On-Demand Loading and Copy-On-Write Brilliance

Tackled advanced process memory management in the xv6 operating system, implementing dynamic on-demand paging and the innovative Copy-On-Write (CoW) Optimization. Features include dynamic On-Demand Binary Loading, a robust Page Fault Handler ensuring continuous process operation, efficient On-Demand Heap Memory loading, and a strategic Page Swapping to Disk mechanism to tackle memory constraints. Enhanced the fork() system call conserving memory by creating separate yet identical spaces for parent-child processes using CoW. Key technologies- C Programming, OS Concepts, Memory Management, Page Fault Handling, and CoW Optimization.

QEMU RISC-V Bootloader Mastery and Enhancing Security and Functionality

Delved into the OS boot process using the QEMU Emulator, creating a custom bootloader for the xv6 OS kernel. Utilized GDB Debugger for Boot ROM inspection and crafted a bootloader using Linker Scripts and Assembler Programming (Assembly). Enabled C code execution through stack setup and introduced user flexibility to boot different OSs. Enhanced security using RISC-V’s PMP Configuration and a custom Secure Boot Implementation. Technologies involved- QEMU, GDB, RISC-V, Assembly, C Programming, Linker Scripts, PMP, and Secure Boot

xv6 ThreadMaster - Dynamic User-Level Threads and Smart Scheduling

Undertook the implementation of user-level threads within xv6 processes by developing a comprehensive User-Level Threading Library (ULTLib). This library allowed for the segmentation of kernel-backed threads into multiple user-level threads with efficient management. Key tasks included ULTLib Initialization, dynamic Thread Creation with precise context management, intelligent Thread Switching and Scheduling incorporating multiple algorithms like round-robin and priority scheduling, and advanced Thread Yielding and Destruction capabilities. Additionally, crafted a new xv6 system call, ctime, tapping into the RISC-V register for accurate timing information. Core technologies and tools encompassed- C Programming, Assembly Language (RISC-V), OS Concepts, Thread Management, and System Calls.

Hybrid Cloud Face Recognition System - Integrating AWS and OpenStack for Enhanced Scalability and Efficiency

Developed a hybrid cloud application utilizing advanced face-recognition techniques with Python. The system processes classroom videos, identifies students’ faces, and retrieves academic records from DynamoDB. Integrated AWS and OpenStack, enabling seamless scalability and cost-effectiveness by utilizing both public and private cloud resources. An extension of Project 2, this approach enhances control and flexibility in triggering AWS Lambda functions through OpenStack VMs.

Cloud-Based Student Recognition and Academic Record Management System

Developed a cloud-based application using Python’s face-recognition library to integrate advanced face-recognition techniques. The system processes classroom videos, identifies students’ faces, and retrieves their academic records from DynamoDB. Implemented on AWS Lambda PaaS for efficient, accurate, and cost-effective student recognition and academic record management.

Image Recognition as Cloud Service

The service will take in images in the PNG format from users, use the deep learning model to perform image recognition, and return the top prediction from the model to the user as plain text. The importance of this service is to provide a scalable and efficient solution for image recognition that can be accessed remotely by users, without the need for them to train their own models or have access to specialized hardware.

End to end learning of steering commands for self driving

Developed a self-driving car simulation using 28k images from cameras mounted on the vehicle, utilizing Udacity’s simulator. Trained a CNN model with multi-task learning to predict angle and throttle based on the simulated images.

Relationship Extraction tackling Large Input Large Output (LILO) problem

Utilized context-aware summarization to improve the performance of relation extraction on Electronic Health Records (EHR) which are known to be long and pose a challenge for state-of-the-art NLP systems. I have Utilized the 2010 Relations Dataset for the research.