Telecom RAG Assistant
● An AI-powered RAG assistant developed to answer telecom domain queries, unifying structured and unstructured data sources into a Chroma vector store, reducing query resolution time to 7 seconds.
● Built an embedding and semantic retrieval pipeline using LangChain to convert user queries to vector embeddings and fetch the top 3 chunks through similarity search.
● Deployed the LLM application using Streamlit, packaging the retrieval and generation components to enable interaction by non-technical users.
AIsle - AI Shopping Assistant
● Built a Streamlit-based shopping assistant using LangChain, Python, and SQLite for natural-language product search, reviews, and customer ratings.
● Defined LLM tools to retrieve relevant products and ratings from SQLite based on user queries, with LLM guardrails to structure results and handle out-of-scope queries without hallucination.
● Implemented image-based product discovery using a vision model to extract image descriptions and retrieve matching products from the SQLite database.
Player Movement Analysis in NFL
● Developed a novel PEI metric using frame-by-frame tracking data of players across 18 weeks of the Super Bowl.
● Integrated schedule error, lateral path error, and radial excess into the final 0-1 efficiency metric to assess player performance in a play and across the season.
● Prepared data-driven presentations for NFL coaches and communicated analytical findings through a well-documented report.
Jersey Citi - On The Move
● Executed data blending in Tableau Prep Builder to logically unify Citi Bike trip data and weather data for the first three quarters of 2025.
● Explored the dependence of trip patterns across different weather variables by creating parameters.
● Retrieved key insights such as membership distribution and top-performing stations across the timeline and conveyed results through building a dashboard and story.
Microsoft Security Incident Prediction
● Built ML models (Random Forest, Gradient Boosted Trees, OvR) on 13M+ records (3.5 GB) using PySpark on GCP, achieving 78.92% test accuracy in predicting security incident severity.
● Engineered features by grouping rare categories, preprocessed high-cardinality columns with nearly 1800 unique values, and encoded 45+ features, improving generalization and model performance.
● Conducted scalability tests on multi-node Spark clusters (3–5 workers), analyzing trade-offs between accuracy, training time, and computational cost.
Customer Insights & Marketing Optimization
● Coordinated with the Data, Analytics & Modeling team to simulate customer targeting for a client using various datasets of orders (20000,14).
● Assessed data quality and completeness in preparation for analysis using MS Excel.
● Analyzed data to target the top 1000 high-value customers based on demographics and attributes.
● Developed an interactive dashboard comprising 8 plots using Tableau to communicate findings and enhance data-driven decision-making.
Bank Loan Case Study
● Preprocessed 0.3 million current and previous loan applications with customer background information and credit scores.
● Analyzed the loan applications across 122 parameters using Jupyter Notebook (Python) and identified potential defaulters using predictive modeling techniques; ensured the bank selectively approved loans to borrowers with a high probability of repayment, minimizing financial risk.
