An AI carbon decision engine for companies in Asia-Pacific. It turns messy operational data into ranked, costed actions. Gemini plans the analysis; a calculation engine produces every number.
Built for the AI Builder Cup 2026 by Google Cloud JAPAC (Sustainability & Social Impact theme) as team Growthr Alpha.
The problem
Companies can measure their carbon footprint, but a carbon report ends at a number. The person who owns that number still has to decide what to do: which actions, in which combination, at what cost, and how sure they can be.
What it does
One loop: measure → diagnose → simulate → act.
- Overview and inventory: the footprint by scope and stream, with the emission factor, source and uncertainty behind every tonne.
- Diagnose: flags the months that moved against their own baseline and traces each spike to the activity behind it.
- Simulate: what-if levers, plus a search that finds every combination of actions that hits a target, like "cut at least 15% at no extra cost".
- Act: actions ranked on carbon, cost, effort and confidence, with a marginal abatement cost curve and an exportable plan.
- Ask CarbonPilot: a Gemini agent that picks from 13 engine tools to answer a question, and shows every tool call it made.
- Your own data: upload a CSV. Rules map the columns they know, Gemini maps the rest, and nothing is calculated until you confirm.
The AI never writes a number
Gemini plans and explains. Every tonne and every dollar comes from the calculation engine: activity × emission factor under the GHG Protocol, with the factor's source attached. A budget guard switches the AI off on its own if spending hits the budget, and the engine keeps serving every page.
Built with
A Laravel web app and a FastAPI engine on Cloud Run, Gemini on Vertex AI, Cloud SQL for PostgreSQL, Secret Manager and Pub/Sub. The demo company is a synthetic 12-month dataset.
