Real Estate IQ: Automated Property Valuation Engine
Business Need
Traditional real estate appraisals rely heavily on subjective judgment and manual comparisons, leading to inconsistent valuations in fast-moving housing markets.
Technical Solution
Engineered an end-to-end regression pipeline comparing standard linear baselines (Ridge, Lasso, Elastic Net) against ensemble algorithms (Random Forest, XGBoost) to model housing values in Sindian District, New Taipei City. Standardized spatial and structural features (transit proximity, amenity density, property age, geographic coordinates) and conducted systematic hyperparameter tuning via GridSearchCV.
While tree-based models outperformed linear baselines on complex interactions, XGBoost was chosen for production due to a lower test RMSE of 6.99 and a narrower train-to-CV error gap (Train 4.67 to CV 7.62), demonstrating significantly stronger generalization stability than Random Forest (Train 3.38 to CV 7.54).
Business Impact
Delivers instant, data-driven price benchmarks to support real estate investors, mortgage underwriters, and property appraisers in making faster, risk-informed pricing decisions.
PayGuard: B2B Invoice Payment Delay Predictor
Business Need
B2B service providers processing hundreds of monthly invoices face recurring cash flow friction due to unpredictable payment delays. The Operations Department required a predictive early-warning mechanism to flag accounts at risk of default, enabling proactive intervention through automated payment reminders, early settlement discounts, or restructured terms before invoices fall into arrears.
Technical Solution
Engineered an end-to-end binary classification pipeline evaluating multiple baseline algorithms (K-Nearest Neighbors, Decision Trees, Logistic Regression) against ensemble architectures (Random Forest Classifier) using historical financial and behavioral indicators.
Data preprocessing included feature scaling, missing value imputation, and robust encoder mapping. Systematic hyperparameter tuning via GridSearchCV identified Random Forest as the optimal production engine, achieving ~90% test accuracy with minimal variance between training and cross-validation scores, proving optimal generalization performance on unseen invoice data.
Business Impact & Feature Drivers
Directly mitigates liquidity risk and reduces administrative overhead. Feature importance analysis revealed that client credit score, debt-to-income ratio, and historical late payment count serve as the strongest predictive signals for predicting timely payment behavior.
