Projects Log
The Work
Every project answers one question: “What does the data actually tell us?” From production ML pipelines to peer-reviewed research — here's the evidence.
Chicago Closure Radar
46× lift in HIGH risk bucket · ROC-AUC 0.807
ML early-warning system that flags Chicago restaurants, cafés, and bookstores at risk of closing — months before they shut down. Built on 312,312 food inspection records and 193,474 business license filings from the City of Chicago's open data portal, zero scraping required. The single strongest signal: how long it's been since the city last showed up to inspect. When inspectors stop visiting, something is usually wrong. XGBoost with SHAP explainability, deployed as a live FastAPI + Next.js app.
Customer Churn Prediction
92% accuracy · Est. 18% churn reduction
Built a production-grade ensemble pipeline on 50,000+ customer records. Engineered 28 features, handled class imbalance with SMOTE, and ran XGBoost alongside Random Forest. Added SHAP so the team could see exactly why the model flagged someone — not just that it did. Estimated 18% churn reduction in simulation.
Influencer ROI & Fake Engagement Detection
IEEE Forthcoming · 8M+ impressions analyzed
Used my own Instagram account @aneela_veldi as the primary research dataset. Applied anomaly detection models to separate authentic engagement from bot-inflated metrics across 4,823 posts. NLP-based comment analysis and time-series engagement modeling produced an interpretable ROI scoring framework applicable to brand marketing decisions in the $21B influencer industry.
Admissions Analytics — DePaul College of Law
20+ hours/week saved · Live dashboards
Replaced a manual Excel-dependent reporting workflow with automated Python pipelines and Tableau dashboards for the admissions office. Built enrollment forecasting models that now drive which prospects to pursue and when to push harder on yield. Automated weekly reports cut 20+ hours of manual work per cycle.
Morningstar Investment Fund Classifier
10K+ fund records · Multi-class classification
Classification system categorizing mutual funds by historical performance, risk-adjusted return metrics, and Morningstar star ratings. Logistic regression and random forest models evaluated across 10,000+ fund records. Combined financial domain knowledge from Whirlpool-era modeling experience with statistical classification methods.
Germany Market Entry Financial Model
5-year DCF · Executive presentation
Feasibility study for a consumer goods client looking at the German market. Covered competitor benchmarking, regulatory gaps, consumer segments, and a 5-year DCF model. Delivered as a PowerBI report with three scenarios — conservative, base, and aggressive — presented to leadership.
ISB Emerging Markets Research
Faculty collaboration · Panel datasets 100K+ records
Research Associate at the Indian School of Business, Asia's #1 business school. Managed panel datasets of 100K+ records, ran econometric models in Stata, R, and Python, and collaborated with faculty on publications studying consumer decisions in South Asian emerging markets.