Work Experience

• Software Engineer Intern @ Amazon.com

  June 2024 - August 2024, Bellevue, WA

  • Implemented an LLM Console to test the REST API Service that powers Alexa LLM.

  • Created a Lambda Function that enables the recovery of logs from CloudWatch to visualize the Computational RexGraph of hosted Reactive Service.

  • Developed a LangChain LLM agent that teaches Reactive Programming and helps developers visualize Reactive Graphs.

• Software Engineer Intern @ Amazon.com

  May 2023 - August 2023, Bellevue, WA

  • Worked with a Data Science team that develops ML models to match user utterances with service providers.

  • Created an Experiment Framework to offload traffic from a particular production ML Model based on a set of eligibility rules.

  • Designed a UI that allows data scientists to create offloading rules that, when triggered, change the incoming traffic of an ML model

  • Improved the time to offload traffic from a model by approx. 30%.

Research

• Web-scale Semantic Product Search With Large Language Models

  E-Bay, Advisor: Dr. Nikhil Muralidhar

  Designed a web-scale semantic product search system leveraging large language models, significantly improving search relevance for e-commerce platforms. Utilized a multi-stage training process with BERT-based models, achieving a 23% improvement over baseline methods while maintaining low inference latency for real-time applications.

• Knowledge-Guided Surrogate Modeling of Subsurface Carbon Plume Evolution for Efficient Carbon Sequestration

  Los Alamos National Lab, Advisor: Dr. Nikhil Muralidhar

  Developed a Convolutional Neural Network (CNN) to model complex, non-linear relationships between gas pressure, gas saturation, and permeability in carbon sequestration. The project utilized a physics-framed deep-learning proxy model to accurately simulate subsurface carbon storage, enhancing the prediction accuracy of carbon capture and storage processes.

Projects

• Synthetic Limit Order Book

  Built using a multi-agent market simulator to model limit order book (LOB) data under varying market conditions. It focuses on the effects of distributional shifts, using shocks in the data to benchmark the robustness of forecasting algorithms. The implementation of the simulator and dataset creation was carried out in C++, providing a controlled environment for testing machine learning models against both in-distribution and out-of-distribution scenarios.

• Batch Processing System

  Architected and implemented a robust batch processing system in C++ to handle high-volume financial data processing tasks, critical for the fund’s trading and risk management operations. Leveraged RabbitMQ for reliable job queuing and asynchronous processing, ensuring efficient handling of large datasets and real-time data feeds.

• Reactive Programming Library for OCaml

  Developed a Reactive Programming Library for OCaml, enabling asynchronous and event-driven programming by supporting the composition and management of data flows. This library simplifies handling dynamic data changes, making it easier to build responsive and scalable applications in OCaml.