Work History

Reproduced and referenced the state-of-the-art ReAct research paper in production to lead the development of an autonomous medical-claim denial AI agent with LangGraph, TypeScript, Java in partnership with one of the largest hospital groups, leveraging thought–action–observation loops for dynamic tool selection and workflow navigation

Developed an agentic chat UI in React and React Query for interactive claim data review and action-taking using Claude

Built a responsive dashboard using Shadcn UI and React to display key metrics such as total claims resolved, and reimbursement amounts received

Fully owned numerous features, including email templating and inbox, to create one of the market's most flexible and powerful AI personalized email sending and monitoring capabilities using TypeScript, Node.js, LangChain, Next.js, React, Chakra UI, Prisma, and PostgreSQL

Implemented caching infrastructure using Redis and worked on data modeling by creating key models and indexing crucial fields to enhance performance and optimize speed using Prisma

Built most of the frontend and wrote end to end tests using playwright to simulate user action and detect user facing bugs

Built order processing automation software for the largest mushroom distributor in Illinois to convert order emails into invoices on QuickBooks using LLM

Developed a tailored accounts payable automation solution for Foxtrot, a company valued at over $500M, which involved extracting information from unstructured PDF documents and organizing it into the ERP system using vector database

Constructed a backend infrastructure to streamline order processing using Express.js and Node.js, while integrating Gmail API, Quickbooks API, and OpenAI API

Raised f1-score from 0.79 to 0.83, precision, and recall for blockchain anomaly detection using PyTorch to provide a more accurate blockchain fraud detection tool for the users

Generated synthetic data on Ethereum transaction data (20×) and benchmarked CT-GAN vs. SMOTE to address class imbalance, improving fraud-detection model performance

Developed a facial alignment and detection algorithm using MediaPipe face mesh and OpenCV to create data pipeline and automate image data preprocessing used for live camera filters in a face-to-face online game

Constructed a CNN Deep learning car model classification algorithm using PyTorch with 90% accuracy in classification of images to suggest corresponding company, model, and year of the car

Extracted 50,000+ image data from Arcane videos using the algorithm to train Deep learning model for image filtering