MoveEasy

Inspiration

Singapore became a super-aged society this year, with 21.4% of citizens aged 65 and above (NPTD, 2026). That share will keep rising, and the burden is local: some planning areas have far more seniors than others, yet planners lack a clear, area-by-area view of where older residents struggle to reach care and community services. We wanted to turn the question "where do we build next?" into an evidence-based answer that anyone can read.

What it does

MoveEasy scores every Singapore planning area from 0 to 100 on how easily seniors reach an ageing centre, a polyclinic and a cycling path, and ranks the 10 areas to fix first, with the gap behind each. A friendly map shows pins, radius rings and colour-coded areas. A what-if simulator lets planners propose a new site or path and see how many seniors it brings within reach. A plain-English chatbot answers questions from the same data. No technical background is needed.

How we built it

We are building MoveEasy end to end on Databricks Free Edition with open data from data.gov.sg.

Data: LTA cycling paths and MRT exits, URA Master Plan 2025 cycling layers and planning area boundaries, MSF ageing centres, MOH eldercare and polyclinic listings, HDB elderly population by planning area, ECDA childcare centres and annual public transport ridership.
Pipeline: a config-driven dataset registry feeds a medallion architecture (bronze, silver with an error table, gold star schema and area features on an H3 grid), governed in Unity Catalog.
ML: a transparent composite accessibility score, a K-means typology and a coverage optimiser, all tracked in MLflow and checked with stability metrics (silhouette, cluster stability, rank stability and top-10 overlap).
Agents: a Genie space for data questions, plus an orchestrator with spatial, simulation and explainer agents on a pay-per-token Foundation Model API.
Delivery: an AI/BI dashboard and one Databricks App with the map, simulator and chatbot.

Challenges we ran into

Scope: nine data sources, a simulator and a chatbot is a lot for two weeks, so we split the build into a core and stretch goals.
Free Edition limits: one 2X-Small SQL warehouse, three apps, five concurrent job tasks and no GPU serving shaped our architecture.
Data gaps: the suggested datasets did not include polyclinics or planning area boundaries, and ridership has no geometry, so we sourced extra datasets and treat ridership as context only.
Evaluation without labels: there is no ground truth for "good accessibility", so we chose unsupervised and stability-based metrics and will be transparent about their limits.

Accomplishments that we're proud of

A complete, buildable architecture that fits Free Edition limits, from raw data to a non-technical user interface.
A star schema, area feature table and ML output tables designed before writing any code.
A measurable evaluation plan with six core metrics, instead of a dashboard with no way to judge its quality.
A clear brand and logo for the project.

What we learned

Accessibility problems are local, so averages hide the areas that matter most.
Define the evaluation metrics before the models, especially for unsupervised work.
Genie only writes SQL, so spatial analysis and simulations need our own agent tools.
Platform limits are design inputs. Scoping early beats cutting features late.

What's next for MoveEasy

Complete the two-week build sprint and rehearse the demo end to end.
Stretch goals: DBSCAN hotspots, Isolation Forest outliers and Unity Catalog governance with lineage.
Overlay the Master Plan 2025 cycling network to show future accessibility gains, and extend the lens to childcare access for working parents.
Reuse the same engine for other domains by swapping datasets in the registry, such as hawker centre access.
Test the rankings and simulator with real planners and ageing centre operators.