Projects
U.S. City Neighborhood Archetypes→
- Developed multi-layer spatiotemporal models integrating 50+ years of census, environmental, and housing data (10+ GB, 500K+ records) using K-Means and GMM clustering to identify 8 distinct neighborhood evolution patterns across 100+ U.S. cities
- Built interactive Mapbox/Deck.gl dashboard forecasting neighborhood changes through 2070, downloaded 2,000+ times by urban planners and researchers, and cited in 2 academic papers analyzing gentrification trends
Customer Churn Prediction→
- Built supervised models (XGBoost, CatBoost, Logistic Regression) on 200K+ telecom customer records achieving 0.89 F1-score to predict churn, and applied K-Means and DBSCAN clustering to segment customers into 6 behavioral groups
- Implemented SHAP interpretability to identify top 10 churn drivers (contract type, support calls, usage patterns), enabling targeted retention campaigns projected to reduce churn by 15% and save $800K+ annually in customer acquisition costs
Economic Connectedness Analysis→
- Applied machine learning and network analysis on 500K+ mobility records and income data across 100+ U.S. regions to quantify social capital disparities, revealing that low-income neighborhoods had 40% fewer cross-class connections than affluent areas
- Designed 5 interactive Python and Plotly visualizations illustrating economic connectedness patterns, presented to 3 policy research organizations to inform community investment strategies targeting $10M+ in social mobility programs
Fine-Tuning Mistral for Prompt Evaluation
- Fine-tuned Mistral 7B LLM on 10K+ prompt-response pairs using RLHF-style evaluation pipelines to automate prompt quality scoring, achieving 85% agreement with human evaluators and reducing manual review time by 70%
- Deployed automated evaluation system processing 1,000+ AI responses daily, enabling rapid iteration on prompt engineering and improving average response quality scores by 25% across production applications
Big Data Analytics for Chicago Crime Patterns→
- Processed 7M+ crime records spanning 20 years using PySpark and SQL, identifying that 60% of violent crimes occurred in just 15% of neighborhoods and peak crime hours clustered between 8PM-2AM on weekends
- Built predictive models achieving 78% accuracy for high-risk areas and deployed Tableau dashboards used by Chicago PD analysts to optimize patrol allocation, contributing to 12% reduction in response times in targeted zones
Graph Neural Networks for Recommender Systems→
- Compared traditional recommender baselines with state-of-the-art GNN models (ConsisRec, DMGCF, KGNN-LS) on large-scale datasets
- Improved recommendation accuracy by 7% and reduced error by 4% using graph-based user–item modeling techniques
Open-Source Contribution – GitHub →
- Collaborated with international developer to enhance Python command-line correction tool with 5K+ GitHub stars, implementing cross-platform compatibility fixes and automated testing suite covering 95% code coverage
- Reduced bug reports by 40% and improved tool reliability across Windows, Mac, and Linux environments, benefiting 10K+ active users and earning contributor recognition in project documentation
BBC News Network Analysis→
- Collected and analyzed 50K+ Twitter interactions using Twitter API to map network of 5,000+ users engaging with BBC News content, identifying 8 distinct information clusters and measuring network centrality metrics
- Uncovered coordinated information spread patterns related to Ethiopian civil conflict with 300+ highly connected accounts driving 60% of engagement, presenting findings that informed BBC's understanding of audience segmentation and misinformation propagation