Data science student and building Trustmebro to make online claims easier to varify

Kimmoji Raja  • Peerlist

I conducted a data-driven study to understand how branding affects customer satisfaction in India’s fast-growing e-commerce market. The project involved analyzing survey responses from 608 users across India and applying machine learning techniques to predict satisfaction levels. Key Contributions: 1. Performed data cleaning, preprocessing, and feature engineering on consumer behavior data. 2. Conducted detailed EDA to uncover patterns in brand loyalty, trust, and platform preferences. 3. Built multiple ML models (Logistic Regression, Decision Tree, Random Forest, XGBoost, CatBoost). 4. Achieved 72.13% accuracy with Random Forest — the highest among all models. 5. Identified top branding factors: customer reviews, product quality, social media presence, and secure transactions. 6. Provided actionable insights for improving customer engagement and loyalty.

The Impact of branding on India's E-Commerce

Developed a computer vision system to detect crop pests in real time and support precision agriculture by enabling timely pest identification and targeted pesticide application. The project aimed to reduce excessive pesticide usage while helping improve crop health through early detection. I collected and prepared annotated image datasets, trained a YOLOv8 object detection model, and evaluated its performance across multiple pest classes. The system identifies pests directly from field images and estimates infestation severity, which can be used to recommend appropriate pesticide dosages based on the affected crop area. The solution demonstrates how deep learning can assist farmers in making faster and more informed pest management decisions. Achieved 99.5% mAP@50 for Bollworm detection, with 100% Precision and 98% Recall. Attained 96% mAP@50 for Stem Borer detection and 91% mAP@50 for Brown Planthopper detection. Successfully detected five different agricultural pest classes using a single YOLOv8 model. Implemented pest severity estimation to generate adaptive pesticide dosage recommendations based on crop area and infestation level. Built a complete pipeline covering dataset preparation, annotation, model training, evaluation, and real-time inference.

Drone Based Multi-Pest Detection for Precisio

Developed a deep learning-based sales forecasting system to predict future product demand and support data-driven inventory planning. The project focused on learning complex sales patterns from historical retail data, enabling more accurate demand forecasting and helping businesses optimize stock management and operational planning. I designed an end-to-end forecasting pipeline that included data preprocessing, feature engineering, model development, evaluation, and prediction. A deep neural network incorporating product and store embeddings was trained to capture relationships between categorical and numerical features, allowing the model to learn purchasing patterns across different stores and products. The solution was evaluated against traditional machine learning models to assess forecasting performance. Project Highlights: Developed a deep learning forecasting model using product and store embeddings to improve demand prediction accuracy. Trained and evaluated the model on a retail dataset containing 150,150 sales records. Achieved an R² score of 0.9771 with a Mean Absolute Error (MAE) of 1.96, demonstrating high predictive performance. Compared the deep learning model with traditional machine learning approaches, including Random Forest and XGBoost, to validate performance improvements. Built a complete forecasting pipeline covering data preprocessing, feature engineering, model training, evaluation, and prediction.

Embedding-Based Deep Neural Networks in Sales

FinSenseAI is an intelligent financial transaction categorization system designed to automate the classification of banking transactions into meaningful spending categories. The project combines Natural Language Processing (NLP) with machine learning to reduce manual effort and improve the accuracy of financial data organization. I fine-tuned a DistilBERT model to understand transaction descriptions and built a complete end-to-end pipeline for data preprocessing, model training, evaluation, and deployment. The application includes an interactive Streamlit interface that supports both individual transaction prediction and bulk CSV processing, along with visual analytics to help users understand their spending patterns. Key Results: Fine-tuned a DistilBERT model on a dataset of 12,000+ financial transaction records. Achieved 97% classification accuracy in categorizing transaction descriptions across multiple spending categories. Developed a Streamlit application supporting both single transaction prediction and bulk CSV classification. Built a complete NLP pipeline covering data preprocessing, feature preparation, model training, evaluation, and deployment. Improved transaction organization by significantly reducing the manual effort required for financial record categorization.

FinSenseAI - Financial Transaction Categoriza

Developed a machine learning-based fraud detection system to identify suspicious financial transactions in real time. The project focused on handling highly imbalanced transaction data by applying advanced preprocessing techniques and training predictive models capable of distinguishing fraudulent activities from legitimate transactions. I engineered transaction-level features, performed exploratory data analysis, and trained a LightGBM model to improve fraud detection performance. To enhance transparency, I integrated SHAP-based explainability, allowing the model's predictions to be interpreted through feature importance. A real-time transaction simulation pipeline was also implemented to demonstrate how incoming transactions could be analyzed with minimal latency. Project Highlights: Developed a fraud detection model using LightGBM, achieving a ROC-AUC score of 0.97 and an F1-score of 0.86 on highly imbalanced transaction data. Applied SMOTE and feature engineering techniques to improve fraud detection performance and reduce the impact of class imbalance. Engineered transaction-level features, including spending behavior, device usage, transaction frequency, and temporal patterns, to improve predictive accuracy. Integrated SHAP (SHapley Additive Explanations) to provide interpretable predictions and identify the most influential features behind each fraud decision. Implemented a real-time transaction scoring pipeline with an average prediction latency of approximately 0.2 seconds, enabling near real-time fraud detection.