More data science, analytics project.

Broadcast overlays need fast, pitcher-specific pitch type predictions that can support real-time on-air usage.

Implemented a scikit-learn pipeline with imputation, scaling, and one-hot encoding, then tuned a KNN model with stratified cross-validation and grid search on a time-based 2025 holdout set.

Achieved 0.985 accuracy and used a normalized confusion matrix to identify class-specific errors and define conservative deployment rules for broadcast use.

Credit risk monitoring needed a structured way to flag overheating markets and prioritize deeper review at the country level.

Conducted rigorous EDA and data QA by removing aggregate regions, resolving valuation-method inconsistencies, and validating long-run credit dynamics across sectors before model development.

Produced country-level risk scoring and stakeholder-friendly visual outputs, including a country risk map, with model performance reaching an AUC of 0.69.

Housing market analysis needed a faster way to compare price trends across states and metro areas without manual filtering.

Built a clean analysis dataset with tidyverse, engineered time features, standardized state and metro identifiers, and implemented reactive UI filters for state, metro area, and year range.

Enabled reliable market trend comparisons with linked time-series plots and a drilldown table for validation in a single dashboard.

The datathon problem required interval-level forecasts that could support staffing, scheduling, and contact center operations planning.

Cleaned and merged multi-sheet Excel and CSV operational data, engineered calendar, lag, and rolling-window features, and built regularized scikit-learn regression pipelines with imputation, scaling, and target transformation.

Generated submission-ready interval forecasts and summarized results for operations planning in a 36-hour datathon setting.