Sai Yashwanth Brahmadevara

Research Assistant

The research focuses on conducting ensemble sensitivity analysis on climate data—specifically precipitation, mean sea level pressure, and geopotential height across the U.S.—to analyze trends in snowstorms and cyclones.

• – Worked on processing climate datasets from GRIB files, generating detailed visualizations and multi-day forecasts. Automated image generation for each climate parameter across a 9-day forecast window to support extensive analysis.
• – Optimized the analysis using multiprocessing techniques to enable parallelization, improving the efficiency of multiple linear regression calculations by running them concurrently with NumPy, ultimately reducing processing time by 65%.
• – Created an interactive web tool that allows users to select specific geographic regions on a world map, automatically tailoring ESA to user-defined latitudes and longitudes.

Software Engineer

As a Backend Developer on Paytm’s Core Settlement Team in the Lending vertical(first few months as intern), my major responsibility is to ensure that the settlement of repayments made by users to clear their loans is processed correctly and efficiently.

Key Tecnologies : Spring Boot, MySQL, Kafka, Kibana, Grafana, Spring Transactions, Locks, Multithreading, JPA.

• – Worked on a large-scale revamp of the postpaid (credit card) system for approximately 4.5 million accounts, implementing validation and correction of transactions using microservices architecture and multithreading for parallelization, as well as Kafka for reliability and scalability.
• – Optimized CSV file processing (approximately 2GB), reducing processing time by 65% through efficient memory management while reading the CSV file and using multithreading.
• – Reduced database calls by implementing synchronized Spring EhCache to cache frequently accessed data.
• – Worked on designing complex SQL queries by combining multiple tables and using row_number(), rank(), CTEs, coalesce, IF, CASE, JOINs, and GROUP BY for large datasets as part of revamp project.
• – Worked on developing late fee waivers for both posting to users and settling them in the repayment process. Also worked on the calculation of days past due.

In addition, worked by following best software practices, such as implementing the Factory pattern, leveraging abstract classes and interfaces, handling exceptions, logging, and creating unit tests using Mockito. Also debugged issues and implemented multiple bug fixes.

Software Developer Intern

• – As a Full Stack Developer, worked on building and maintaining a dynamic dashboard for multi-service data display, using AngularJS (Angular Material) on the frontend and Spring Boot with microservices on the backend.
• – Enhanced database performance through efficient SQL query construction, index optimization, and table partitioning.
• – Developed and optimized CI/CD pipelines with GitHub Actions and AWS CodeDeploy, enabling seamless deployments to the production environment.

Stony Brook University

Teaching Assistant - CSE 214 Data Structures (Java Programming) Prof. Ahmed Esmaili

GPA: 3.72

Compiler Construction

Worked on building a compiler for a custom architecture that runs a BERT model on bare metal. Developed custom passes in MLIR and LLVM to generate the necessary instructions, converting the PyTorch BERT model to the TOSA dialect in MLIR. Created a custom dialect to lower TOSA operations to the target dialect. Designed new LLVM intrinsics to support tensor-specific operations such as softmax and matrix multiplication directly. Extended RISC-V with additional instructions for tensor operations, including softmax, layer normalization, and matrix multiplication. Defined new register types to handle tensor pointers and shape metadata as "tensor registers."

xv6 : Implemented Demand Paging and Software TLB

Designed and implemented a software-managed Translation Lookaside Buffer (TLB) between hardware and the OS kernel to optimize performance by minimizing page table lookups and improving memory access times. To further enhance efficiency, I developed an application-controlled page eviction strategy inspired by the Exokernel architecture, exposing virtual-to-physical address mappings directly to applications for finer control over memory management. This project involved creating new system calls, managing hardware interrupts, and integrating low-level mechanisms to ensure seamless interaction between software and hardware layers.

Comparison among distributed key value pair databases

Explored distributed key-value stores, including Cassandra, Apache Ignite, and TiKV, to study consensus algorithms such as Paxos, Raft, and Sharding, analyzing their design and use cases. Tested these databases on YCSB benchmarks to evaluate their performance and scalability characteristics.

End to End Movie Booking App

Developed a mobile application frontend with React Native, implementing REST API calls to the backend using Axios. Built an animated ImageSlider using useEffect and React Native’s animation library, and utilized native components such as Navigators, FlatList, and Image. Designed and developed the backend with Spring Boot and MySQL, leveraging Java Streams, Spring Data JPA, and implementing core functionalities.

Machine Learning & Deep Learning

Developed a new variant of the Vision Transformer (ViT) by adding convolutional layers to the standard ViT model and tested it on the MNIST dataset. Built and tested standard and pre-activation bottleneck variants of ResNet on CIFAR-10 using PyTorch. Also implemented standard RNNs, attention models, PCA, and autoencoders as part of coursework.

Conducted stock price prediction for Microsoft using traditional time series models, including ARIMA, seasonal ARIMA (S-ARIMA), and LSTM. Enhanced the prediction accuracy by incorporating additional features derived from financial data and sentiment analysis of Twitter data.