Mayank Kantharia

Delivered a Shiny-based analytics experience that turns raw economic data into navigable, stakeholder-friendly insights.

• Built an interactive dashboard analyzing global economic trends with drill-down regional and sector exploration
• Designed clear, accessible visualizations for non-technical stakeholders
• Structured and transformed raw datasets for efficient, user-driven reporting and analysis
• Designed geospatial maps and time-series visualizations with drill-down country profiles
• Cleaned, standardized, and merged multi-source datasets to ensure consistent indicators

Deployed a full-stack music subscription platform on AWS, combining compute, storage, serverless APIs, and NoSQL data design.

• Deployed an EC2 Ubuntu server with Apache2 and public DNS access
• Designed DynamoDB schemas for login, music, and subscription data
• Built backend logic for registration, subscription management, and music queries
• Integrated API Gateway and Lambda for REST-based workflows
• Automated artist image retrieval and secure storage in S3

Analyzed cross-platform discourse on Reddit and YouTube using NLP, topic modeling, and network science to study narrative spread.

• Collected and preprocessed 90,000+ Reddit and YouTube comments
• Built an NLP pipeline with VADER sentiment analysis and LDA topic modeling
• Detected platform-level divergence (YouTube Q=0.895, 600 communities vs Reddit Q=0.571, 237 communities)
• Modeled information diffusion with the Independent Cascade algorithm
• Identified centralized YouTube hubs amplifying leader-focused narratives

Built and deployed NLP classification pipelines and a Flask web app for real-time review-based recommendations.

• Engineered five feature representations (BoW, TF-IDF, embeddings, title-only, combined)
• Title+Review features achieved 88.97% mean accuracy across 5-fold CV
• Benchmarked Logistic Regression, FastText, and deep learning models
• Combined features outperformed single-input baselines by up to 2.3%
• Deployed the best model in a Flask app with prediction, search, and review submission

Advanced few-shot hyperspectral image classification using attention and temperature scaling to improve generalization.

• Beat published benchmarks with 99.55% overall accuracy vs 98.16% baseline
• Validated attention as the key driver via ablation studies
• Temperature scaling alone underperformed the baseline at 97.21%
• Performed spectral-spatial preprocessing and feature extraction
• Evaluated across multiple HSI benchmarks for generalization

Built an end-to-end churn prediction pipeline on IBM Telco data with modeling insights tied to retention actions.

• Built an end-to-end churn pipeline on IBM Telco data (7,043 rows, 21 features) to surface key churn drivers
• Handled class imbalance with stratified sampling (5,000 records) and stratified CV
• Set up preprocessing with scaling, one-hot encoding, and ANOVA F-test feature selection
• Reached 0.836 ROC-AUC with Logistic Regression across five ML models and one deep learning model
• Produced a reproducible analysis with separate EDA and modeling notebooks plus model comparison tables