Population-Health
Intelligence Platform
Calibrated life-expectancy forecasts for 193 nations — with the signals driving each trajectory.
01 · Problem
Planning decisions deserve more than a point estimate.
Public-health planners decide on top of life-expectancy estimates. The numbers they get are national averages with no signal about why a trajectory is shifting.
Two countries can decelerate for opposite reasons — declining immunization, GDP contraction — and need different interventions. The output should reflect that.
Point estimates only
No signal on why
Wrong interventions
02 · Solution architecture
From indicators to a planner console.
Validated features in, calibrated forecast out — with attribution.
The spine
Sources
daily ingest
Feature Store
versioned
Forecaster
ensemble + quantile
Inference API
/predict /attribute
Sources
WHO, World Bank, IMF. Daily ingest, schema validated.
Feature store
Versioned features keyed by country × year × indicator.
Forecaster
Gradient boosting with quantile regression for CI bands.
Inference API
/predict, /attribute, /scenario — every response carries CI + SHAP attribution.
03 · Live forecast
Watch the model reason.
One country, one scenario, one forecast — with the signals that drove it. Step through it or let it run.
The country
Baseline · 72.3 years (2015)
The scenario
The forecast
Projected
72.3 years
±1.5 (90% CI)
vs. baseline · +0.0y
1. Country — Pick a country — the surrogate loads its 2015 baseline and feature defaults.
Interactive prototype · deterministic surrogate of the trained ensemble.
04 · Impact
Forecast with reasoning.
The output isn't a number — it's a number with the signals that produced it, calibrated and ready for a planning conversation.
193
countries covered
±1.8y
typical 90% CI band
<50ms
per-scenario inference
05 · What's next
From a forecast to a planning surface.
The same inference layer extends naturally into streaming data, scenario comparison, and drift monitoring.
Streaming ingest
Daily snapshots replaced with WHO/WB change-data-capture.
Scenario diff
Submit two override vectors, get a structured comparison payload.
Drift monitors
Track distribution shift, auto-flag when calibration degrades.
Forecasting that shows its work.
Happy to walk through the calibration layer, the attribution chain, and what a production deployment looks like.