Data Sources & Methodology

Environmental data, then Amazon S3 storage, processing and feature engineering, the Nexus engine (gradient boosting), Amazon SageMaker AI inference, Amazon Bedrock explanation, the AquaHeat agent, and the Flask API with the UI. Green dots are active; yellow dots mean a fallback is in use.

• Nexus score: the model outputs one 0 to 100 score from five features (temperature anomaly, rainfall deficit, water-demand growth, green cover, groundwater stress). Levels: .
• Heat score: 35 + 14 x temperature anomaly + 0.1 x (30 - green cover %), capped at 0-100.
• Water-scarcity score: 30 x rainfall deficit + 30 x demand growth + 40 x groundwater stress, each scaled to 0-1 (deficit -20 to 60%, demand -30 to 50%, groundwater 0 to 100).
• Flood and waterlogging score: rule-based. Storm rainfall (45 mm/day design storm in demo mode, or the live 7-day forecast maximum) x runoff coefficient (0.15 + 0.75 x impervious share) is compared with drainage capacity (70 x (1 - 0.45 x impervious share) mm/day). Score = 100 x (runoff / capacity)^1.3, capped. Imperviousness and drainage are modelled, not surveyed, and no observed flood data is loaded.
• Action plan: each action is applied to the inputs under a stated assumption and the scores are recomputed; benefit is the resulting score reduction. Costs are indicative tiers, not quotes. Actions are ranked by urgency, then benefit per cost.
• City Survival Twin: the districts are a stylised layout of the eight modelled areas, not a surveyed map. Every district value comes from the same model and formulas as the dashboard.
• Community reports: leak, tanker, shortage and standing-water reports are stored as COMMUNITY data. The sample reports shown at first are labelled SAMPLE and counted separately. Consumption anomalies use a median and MAD robust z-score on readings you provide.
• Why (explainability): each factor is reset to its seasonal baseline and the drop in score is reported as that factor's points. These are model attributions, not causal proof.
• Confidence: 12 models trained on resampled data predict under 20 random perturbations of the inputs (240 draws). Confidence is the share of draws that land in the same risk level; the likely range is the 5th to 95th percentile. The input-uncertainty sizes are assumptions.
• What-If: the same model re-scores the scenario. The rainfall slider is a percent change in rainfall (drier means higher deficit and higher risk).
• 30-day projection: a scenario-based projection from the model, not a weather forecast.

• The model is trained on simulated environmental-stress data. Hold-out scores are measured against the same simulator, so they show the model learned the simulation, not real-world accuracy.
• Live mode uses real Open-Meteo temperature, rainfall and soil moisture. Vegetation is a soil-moisture proxy, not NDVI. Water demand is anchored to press-reported Gurugram figures (not meter data) for Gurugram-area locations and modelled elsewhere; no open live water-demand feed exists for NCR.
• Roadmap: train on historical IMD, CGWB and satellite NDVI data, validate against observed shortages, and add utility meter data.