Samit S Email & Phone Number

Wilmington, Delaware, United States

• Work with senior leaders, business units, and partners to influence business and technology roadmaps• Set up AI/ML function for First Party Fraud detection. POC for portfolio-level credit abuse ML model• Developed 3000 model features using python and hive for credit abuse Light GBM model• Apply encoding to categorical features to develop Neural Network model for credit card transaction fraud• Developed and implemented NLP models using deep learning techniques such as LSTM and Transformer to perform sentiment analysis on customer reviews, resulting in improved product development and customer satisfaction.• Developed and implemented linear and logistic regression models to get la customer churn for a telecom company, resulting in a 20% reduction in churn rate and increased revenue.• Also leveraged LLM model API to generate labels and further supply it to supervised models.• Streamline AI/ML function for credit card Third Party Fraud detection. Improved existing model performance by hyperparameter tuning using Optuna – 2% lift• Evaluate model performance using KPI such as Precision, recall, AUC, confusion matrix, ROC curve• Sentiment analysis on U.S. Bank mobile app reviews using Hugging Face Transformer-based model BERT and pipeline• Applied K-means clustering for document segmentation and Topic Identification to understand customer complaints• Conducted text preprocessing and feature extraction using NLP techniques such as tokenization, stemming, and word embedding to prepare data for modeling, resulting in improved model accuracy and reduced overfitting.• Conducted exploratory data analysis and feature engineering to prepare data for LLM modeling, resulting in improved model accuracy and reduced overfitting.• Named Entity Resolution (NER) to extract merchant name from transaction description of DDA transactions using regex, fuzzy-wuzzy and spaCy python packages• POC to transition ML model training from on-premises servers to AWS SageMaker

Wilmington, Delaware, United States

• Lead a team of 6 in design, development, and production implementation of machine learning models to mitigate: Portfolio Credit Abuse Risk, Credit Card Return Payment Fraud Risk and Acquisition Fraud Risk. Expected annual savings-$200M +. (ML Algorithms: Xgboost, ANN, RNN, Random Forest, Bayesian Optimization, SHAP)• Credit Abuse Fraud:Developed model to detect fraudulent customers post acquisition for Credit CardsDeveloped and implemented Credit Abuse Fraud model leveraging distributed Xgboost package in in SCALAUsed a open source LLM model to to train on own document structured.Model is used in production as a rank ordering tool to detect high risk account, which will be closed after manual reviewPerformed Feature Engineering to generate 25000 variables for model development using Pyspark and SASDeveloped variable selection methodology to reduce initial set of variables for hyperparameter tuningLeveraged Random Search, Grid Search and Bayesian optimization techniques: Hyperopt for hyperparameter tuningLeveraged GPU based environment to do hyperparameter tuningDeveloped new variable reduction methodology to reduce final model variables without compromising model performanceCreated challenger model using Neural Network models: ANN, RNN and Random ForestGenerated model reason codes using SHAP for operation analyst to assist in making decisions on account closurePresented model results and described model use to different business teams, model governance and senior executives• Credit Card Return Payment Fraud:Developed Return Payment Fraud machine learning model to detect high risky credit card paymentsModel implemented in production is used to place hold on high-risk payments and prevent losses due to immediate release of creditDeveloped and implemented Return Payment fraud model leveraging distributed Xgboost package in SCALALeveraged Rule Fit algorithm implementation in R to extract rules from initial Xgboost model

Gurugram, Haryana, India

• Built logistic regression/random forest models in SAS and R to predict fraud claims in worker’s compensation insurance• Generated model hints/reason codes to help insurance analyst to evaluate claims• Applied NLP using R programming to extract themes from eight million unstructured call center texts generated from inbound and outbound call notes• Categorized calls based on call themes/Topic Identification using clustering algorithms, including hierarchical and K-Means algorithms