Inspiration
Singapore is good at reducing waste: daily domestic waste per person has fallen about 21% in a decade. But it is not getting better at recycling. The overall recycling rate sits at 52% (59% in 2019), and the domestic rate is just 11% (17% in 2019), against a 70% Zero Waste Masterplan target for 2030. Total tonnage disposed has even crept up, from 3.04 Mt in 2014 to 3.33 Mt in 2024, with Semakau Landfill's runway finite.
What struck us is that the open data already exists (bin locations, e-waste points, recycling time series), but nobody has joined it into an answer to a planner's real question: where should the next recycling dollar go? That question inspired ZeroWaste Navigator.
What it does
ZeroWaste Navigator is a decision tool, not just a dashboard:
- Maps the gap. It computes recycling-bin coverage per 1,000 residents for every planning area.
- Projects the trend. It forecasts the national recycling rate to 2030 from the historical series.
- Ranks where to act. It surfaces the top 5 planning areas where investment would improve recycling the most.
- Shows before vs. after. It simulates an intervention and compares the projected national rate with and without it.
How we're building it
On Databricks, the pipeline runs: data.gov.sg / NEA sources → Lakeflow ingestion → Delta bronze/silver/gold tables governed by Unity Catalog → SQL spatial joins and MLflow-tracked forecasting and scenario models → a Databricks App with an AI/BI dashboard and Genie for plain-English questions.
What we're learning
- Singapore's recycling story is two stories: non-domestic recycling is around 67%, but households are at 11%, so a national average hides where the real gap is.
- Per-capita waste falling while total disposal rises shows why recycling, not just reduction, is the lever that moves landfill pressure.
- Good planning tools need to be explainable. A simple, auditable gap score is more useful to a planner than an opaque model.
Challenges
- Denominators: Bin counts mean little without residents per planning area, so we have to join population data to geospatial layers with consistent boundaries.
- Bins are not outcomes. Coverage is a proxy for recycling behaviour, so we are careful to frame the scenario model as an estimate with stated assumptions, not a guarantee.
- Short time series. Annual national data gives few points for forecasting, so we favour simple, interpretable models with uncertainty ranges over complex ones.
- Scope. Route optimisation, food-waste estimation from hawker centres and behaviour modelling are tempting, but we are prioritising a finished core demo over stretch goals.
What's next
A working Databricks App with the coverage map, the top-5 ranking and the before-vs-after comparison, and then the stretch goals if time allows.
Log in or sign up for Devpost to join the conversation.