Nithya Sharma

Project Overview

This project focused on developing a machine learning-based system for detecting fraudulent financial transactions in real-time. The solution helps financial institutions minimize losses and protect customers from unauthorized transactions.

Technical Approach

• Preprocessed and explored a dataset of anonymized financial transactions, handling missing values and outliers.
• Addressed severe class imbalance (fraud cases were less than 0.2% of the dataset) using SMOTE (Synthetic Minority Over-sampling Technique).
• Performed feature engineering to extract transaction patterns and temporal behaviors.
• Implemented a comparative analysis of multiple machine learning algorithms including Logistic Regression, Random Forest, and XGBoost.
• Optimized hyperparameters using GridSearchCV with stratified k-fold cross-validation.
• Fine-tuned model for precision to minimize false positives while maintaining high recall for fraud detection.

Results & Impact

The implemented solution was estimated to reduce financial losses by up to 87% by identifying fraudulent transactions before they were processed. The false positive rate was kept below 0.05%, ensuring minimal disruption to legitimate customer transactions.

Technologies Used

Project Overview

This project aimed to predict employee attrition risks by analyzing various HR factors. The goal was to enable organizations to implement proactive retention strategies to improve employee satisfaction and reduce turnover costs.

Technical Approach

• Analyzed a comprehensive HR dataset containing employee information including satisfaction levels, performance evaluations, project counts, average monthly hours, and tenure.
• Performed exploratory data analysis to identify patterns and relationships between variables.
• Created new features to capture workload imbalance, overtime patterns, and satisfaction-to-compensation ratios.
• Implemented and compared multiple classification models: Logistic Regression, Decision Trees, Random Forest, and Gradient Boosting.
• Used cross-validation to assess model performance and prevent overfitting.
• Applied feature importance analysis to identify key factors influencing employee attrition.

Key Insights

The analysis revealed several significant factors contributing to employee attrition:

Results & Implementation

The Random Forest classifier produced the best results with 94.7% accuracy. The model was implemented as part of an HR dashboard that provides:

• Real-time attrition risk scores for individual employees
• Department-level risk assessments
• Specific recommendations for intervention based on the key factors identified for each high-risk employee
• Projected financial impact of potential turnover

Technologies Used

Project Overview

This project developed a real-time phishing detection engine that analyzes URLs to identify potentially malicious websites. The solution was designed to be lightweight enough for browser extension integration while maintaining high accuracy.

Technical Approach

• Collected and preprocessed a dataset of 11,000+ URLs with binary labels (legitimate or phishing).
• Extracted 30+ features from each URL, including:
  • URL length, domain age, and presence of HTTPS
  • Domain reputation and registration information
  • HTML and JavaScript characteristics (e.g., iframes, redirect count)
  • Form submission targets and credential input detection
  • Link destination analysis and anchor tag behavior
• Applied feature selection techniques to identify the most predictive features.
• Implemented multiple classification algorithms and evaluated their performance.
• Optimized the final Gradient Boosting Classifier model for both accuracy and inference speed.

Implementation & Deployment

The model was deployed in two complementary forms:

Results & Performance Metrics

In real-world testing, the system successfully identified 94.3% of previously unseen phishing sites within the first 24 hours of their appearance online, providing crucial early protection for users.

Technologies Used