Selected portfolio projects with full narrative — business context, methodology, results, trade-offs, and deployment sketches. Each case study draws plots from its executed notebook so the page stands alone.
Day-ahead and short-horizon load forecasts on 16 years of real PJME hourly consumption — comparing SARIMA, state-space (UnobservedComponents + Fourier exog), and gradient-boosted forecasters. The "best" model depends on whether procurement or risk teams are reading.
Daily lake-level forecasting for the world's largest man-made reservoir, on real Zambezi River Authority data. A 7 cm RMSE on a 7 m operational band turns "are we going to breach turbine safety thresholds in the next 30 days?" from a guess into a calibrated probability.
10 years of daily NASA POWER data over Nairobi, three forecasters compared against a stubbornly hard baseline: monthly climatology. The headline finding is uncomfortable for forecasters — at this latitude the seasonal envelope is most of the signal, and same-day weather covariates are what actually move point accuracy.
Three GLM families and a gradient-boosted challenger compared on the canonical French motor third-party liability dataset (678k policies). Tweedie wins on segmentation power (Gini); Poisson + Gamma wins on top-decile lift. Picking between them is an actuarial decision, not a modelling one.
Mobile-clinic dispatch as a Markov decision process on real KMPDC + SHA Kenyan data. Three policies — manual round-robin, capped LP, tabular Q-learning — compared on patients-served, travel cost, and equity. The honest finding: the algorithm matters less than how the constraints are written.
