HeatGuard NightWatch

Inspiration

Singapore becomes super-aged in 2026, and 2026 is also the Year of Climate Adaptation. Those two facts collide in older HDB rental flats, where many seniors live alone without air-conditioning.

What struck us was a 2026 NTU - TTSH - MOH study: nights above ~25.1°C raise emergency visits, with effects sometimes appearing weeks later. Yet today's heat response is daytime and island-wide, and cooling centres only open once a heatwave is forecast. In January 2026, Parliament even asked whether cooling spaces could be prioritised for seniors living alone. Nobody can currently tell outreach teams which blocks to check before a heatwave is declared. That gap is what we want to close.

What it does

HeatGuard NightWatch scores heat risk for ~13,000 HDB blocks every 15 minutes and turns it into a nightly check-in list for Active Ageing Centre staff.

Cumulative heat load, day and night: consecutive hot days and warm nights, not just the 3pm peak
Block-level vulnerability: elderly density, 1 - 2 room rental flats, building age and distance to the nearest AAC or community club
Action, not just a map: a Databricks App gives each AAC its top-10 blocks, the nearest cooling space, and a check-in script in English, Chinese, Malay and Tamil
For planners: a what-if view ("open a cooling space here") and a Genie space for plain-English questions

The risk score combines exposure and vulnerability:

How we'll build it

All on Databricks Free Edition, using open data from data.gov.sg (NEA WBGT, temperature and rainfall APIs; HDB property and elderly data; MSF centres; PA community clubs):

Bronze: a Lakeflow Job ingests 3 live NEA APIs every 15 minutes into Delta tables
Silver: a Lakeflow Declarative Pipeline with data-quality expectations cleans readings and maps each block to its nearest station and AAC
Gold: rolling heat-load and vulnerability features per block
ML: a gradient-boosted classifier for next-24h Low/Med/High risk, tracked in MLflow and registered in Unity Catalog, which must beat a "tomorrow = today" baseline
Serve: a Databricks App, AI/BI dashboard, Genie space, and Foundation Model API for the multilingual briefs
Govern: Unity Catalog lineage from API to score

Challenges we expect

Sparse sensors: only 15 WBGT stations, so block-level heat has to be estimated between stations
Short history: WBGT data only goes back to Feb 2025, about two hot seasons for training
No outcome labels: heat-illness case data isn't public, so we validate against NEA heat-stress levels instead of hospital admissions
Free Edition quotas: one pipeline and limited compute, so we design small and keep a cached fallback for the live demo

What we learned

Researching this changed how we framed the problem. Heat risk isn't just about the hottest hour: it builds up over consecutive warm nights and lands hardest on people the current system can't see. The most useful output isn't a prettier heat map. It's a short, actionable list someone can call tonight.