Machine learning · 2026
FloodCast Jakarta
Hyperlocal flood-risk early warning for 15 Jakarta neighbourhoods: calibrated multi-horizon XGBoost with SHAP explanations in Bahasa Indonesia and an offline map dashboard.
- forecast horizons — 6/12/24 h
- 3
- pilot neighbourhoods
- 15
- recall constraint on alerts
- ≥70%
Architecture
Select a stage to explore it.
01 · Data sources
Three real-world sources fused: BMKG rainfall, Jakarta floodgate telemetry, and digital elevation terrain data for the Ciliwung corridor.
Visuals

01 · Problem
Jakarta floods every rainy season, and the Ciliwung corridor is among the hardest hit — yet early warning is district-level and often too late for neighbourhood evacuation. Responders need probabilities per kelurahan, hours ahead.
02 · Solution
A multi-horizon prediction system that outputs calibrated flood probabilities at 6, 12, and 24 hours for 15 pilot neighbourhoods, classifies them into the four operational alert levels responders already use (Aman → Awas), and explains every prediction in plain Bahasa Indonesia.
03 · Engineering decisions
Flood events occur in only ~2–8% of hours, so imbalance is handled explicitly: weighted training plus a threshold sweep that maximizes F1 subject to a minimum 70% recall — because in early warning, a missed flood costs more than a false alarm. The dashboard runs fully offline by design; infrastructure that fails during the disaster it predicts is decoration.
04 · Results
An end-to-end system built under datathon constraints: real messy data from three sources, calibrated multi-horizon models, SHAP-backed advisories for three audiences, and a map dashboard that runs from a single local command.
05 · Lessons learned
Model metrics have operational meaning: choosing a recall floor is a policy decision expressed in code. Competition deadlines don't excuse unstructured code — the modular package made every late-stage change cheaper.