Inspiration

Singapore is getting hotter and hazier. Heat matched the country's all-time record of 37.0°C in 2023, with 37 days of high heat stress that year. On 15 September this year, haze pushed all five regions into unhealthy air quality for the first time since 2019. We went up to Marina Bay Sands in September and couldn't see the skyline, and people around us were falling sick.

What struck us is that the warning is the same for everyone: a single number and a generic advisory. A fit 25-year-old and a 78-year-old with COPD get the same message, even though seniors, young children and outdoor workers face far more risk. Heat and haze hit the same vulnerable groups, so we wanted one system that protects them from both.

What it does

ShieldSG is a hazard-agnostic early-warning platform. Every hazard follows the same pipeline:

Signal → Forecast → Who's exposed → What to do

Signal: temperature, heat stress (WBGT), PSI/PM2.5, wind, fire hotspots and rainfall from open data
Forecast: a model predicts the risk level 1-3 days ahead
Who's exposed: the forecast is joined with demographics (seniors, young children, eldercare centres, schools) to give a risk score per planning area
What to do: an AI/BI dashboard for agencies and a natural-language assistant for the public, giving tailored advice such as "Is it safe for my father with asthma to take the bus to the clinic tomorrow?"

Two use cases:

Heat (problem statement B1, HeatGuard): maps heat-stress risk to vulnerable communities by planning area.
Haze (live in Singapore now): forecasts PSI from fire hotspots and wind direction, then maps air-quality risk to the same groups.

Flash floods reuse the same pipeline, so adding a new hazard means adding a data source, not rebuilding the product.

How we plan to build it

Ingest open data from data.gov.sg, NASA FIRMS and SingStat into Delta tables governed by Unity Catalog
Train and track the heat-stress and haze forecast models with MLflow
Build the exposure layer by joining forecasts with vulnerable-population data by planning area
Deliver it through an AI/BI dashboard, a Genie-based assistant and a Databricks App with a risk map

Challenges we anticipate

Getting enough historical fire, PSI and heat-stress data to train reliable forecasts
Matching hazard readings from a few stations to planning-area demographics
Giving useful health guidance without presenting it as medical advice

Who benefits

Seniors, people with asthma or COPD, young children and their schools, outdoor workers, and the agencies that need to target alerts early.

What's next

In the 2-week sprint: week 1, build the heat and haze data pipeline and forecast models; week 2, deliver the risk map, a Genie-based advisor and a Databricks App. Flash floods are a stretch goal.