Five projects demonstrating the complete policy intelligence pipeline: from satellite measurement, through causal inference, to policy transfer and Bayesian decision-making.
Feature extraction from satellite data (Sentinel-2, SMAP, BRIGHT) using computer vision and deep learning. Foundation models create learned representations for ecological similarity and land-use classification.
Embedding-informed donor selection improves causal inference for environmental policy. Shows that embedding-matched controls reduce pre-treatment error 49% (0.45→0.23) and reveal 4.4% wildfire mitigation effect (95% CI: [−5.1%, −3.7%]).
Vietnam's Alternate Wet Drying (AWD) water-saving policy evaluated for transfer to Japan. Water balance modeling + biophysical constraints reveal AWD feasible in 18% of Japan (vs. 45% Vietnam), concentrated in Kanto/Tohoku. Spatial fragmentation raises extension costs 3.8×.
Vietnam's 23–28% anti-dumping tariffs on Chinese Hot-Rolled Coil steel. Firm-level analysis reveals policy incidence is inverted: domestic mills profit (duopoly control) while downstream factories lose (24% cost increase). 35% tariff exemptions create institutional capture; macro outcome: stagflation risk.
Which model best predicts Argentina's government wage spending response to tax revenue? Bayesian comparison across 6 models. Student-T cubic wins decisively (ELPD 289 vs. Normal Linear 125)—200 million times more likely. Heavy tails capture outliers; cubic captures non-linear thresholds.