Pradyumna Agarwal

Electricity Consumption Forecasting (Time Series)

Objective: Forecast electricity consumption using PJM hourly data and analyze seasonal and daily usage patterns. Approach: Performed EDA, time-based feature engineering, and built models including ARIMA, SARIMA, Prophet, LSTM, Holt’s, and XGBoost; evaluated using RMSE and MAE. Business Need: Enables energy providers to optimize demand planning, manage grid load, and reduce operational costs. Impact: LSTM delivered best performance; forecasted next 30 days and identified key consumption patterns.

Patient Condition Classification using Drug Reviews (NLP) ⁠

Objective: Analyze patient drug reviews to classify medical conditions (Depression, High Blood Pressure, Type 2 Diabetes) and understand patient sentiment and drug effectiveness. Approach: Performed text preprocessing (cleaning, stopword removal), feature extraction (TF-IDF), and built models including Logistic Regression, SVM, Random Forest, Naive Bayes, and DistilBERT; evaluated using accuracy and F1-score. Business Need: Helps healthcare providers and pharma platforms understand patient feedback, improve treatment decisions, and recommend suitable medications. Impact: DistilBERT achieved best performance; accurately classified conditions and provided insights into drug effectiveness and patient sentiment.

Alzheimer’s Disease Prediction using Medical Data

Objective: Analyze health, lifestyle, and medical data to predict the likelihood of Alzheimer’s disease and identify key risk factors for early detection. Approach: Performed data preprocessing, feature selection, and built classification models including Decision Tree, Random Forest, SVM, Neural Networks (MLP), and XGBoost; evaluated using accuracy, precision, and recall. Business Need: Supports early diagnosis and risk assessment, enabling healthcare professionals to improve patient management and decision-making. Impact: XGBoost achieved best performance; identified key risk factors and enabled accurate prediction of Alzheimer’s likelihood.

PDF Conversational Chatbot using RAG

Objective: Developed a conversational AI chatbot that enables users to upload PDFs and interact with their document content through natural-language questions and follow-up queries. Approach: Implemented a RAG pipeline to extract and chunk PDF text, generate embeddings using all-MiniLM-L6-v2, retrieve relevant content using cosine similarity, and generate grounded responses with Google Gemini; incorporated session-based memory for multi-turn conversations. Business Need: Enables users to quickly retrieve and understand information from lengthy documents through conversational, context-aware interactions. Impact: Built and deployed a functional conversational document assistant with grounded responses and follow-up conversation capability.