- Designed a scalable, real-time anomaly detection model using autoencoders to identify fraudulent coupons, with MLflow tracking and retraining to address drift.- Engineered a Retrieval-Augmented Generation (RAG) model with LangChain, integrating Snowflake and Faiss to enable SQL generation for non-technical users.- Led end-to-end deployment of a predictive survival model, reducing churn through weekly batch predictions and personalized email campaigns.- Built production-grade ensembled ML models for high-value customer acquisition, achieving sustained performance through quarterly retraining.- Developed a ranking algorithm based on purchase history, segmenting 200,000+ products to enable personalized recommendations.- Conducted deep-dive analyses on payment data, translating insights into actionable recommendations to improve payment success.- Designed A/B tests around ML models, defining strategies and refining campaigns with marketing teams, presenting insights to executives.- Applied causal inference to historical data, identifying trends and preferences, and evaluating vendor offers to secure future funding.- Led technical projects while mentoring junior team members, guiding them on code reviews, version control, and GenAI applications.

• Launched a machine learning model to estimate the optimal premium pricing for insurers which centralized the underwriting process that has been in practice for a decade and made a profit of over 380,000$ in two quarters.• Designed and deployed an end-to-end machine learning model to predict the size of an insurance claim. The direct result of the model is the 28% increment in the productivity of claim adjusters.• Explored the raw unstructured XML sheets using python’s XMLTree and ensembled Lightgbm and Catboost machine learning models to boost the accuracy from 74% to 81%.• Spearheaded the common ml framework to handle the process of launching an ML model with ease and packed it as public python package that can be accessed by 8 different sister companies.• Deployed the model using flask and created a pipeline to connect the guidewire for the real-time data to make predictions and demonstrated the model’s behavior to the VP and CEO visually.• Worked solely on Bi-directional Recurrent Neural Network (LSTM) model which involved using NLP techniques like tokenizing, lemmatizing, tf-idf vectorizing, Word2vec word embedding, and BERT embedding to solve the problem of classifying risk types which resulted in reducing the time customers spent on the page by 30%.• Centralized hundreds of different data collection requests by creating SQL scripts and embedding them into a python app which dramatically expedited the process.

• Developed a very user-friendly web application for non-tech federal employees using R shiny for the filtering, sorting, analyzing, and retrieval of federal agricultural data.• Data collected and aggregated data from multiple federal sources of data like NASS, WASDE, AMS etc into a single structured repository.• Built regression models using linear regression and support vector machines on high-dimensional data to predict product sales.• Created a classification model using the Xgboost algorithm that classifies the products as profitable or non-profitable with an accuracy of around 84 %.• Built a time series model using prophet that accurately estimates the projections of various food product's price change in the future and integrated it into the flask web application.