Vayu
National Air-Quality Intelligence and Enforcement Platform
Captured from the live deployment, not a mockup
01 · The problem
Built the loop a regulator actually needs: a citizen photographs a source, a vision model verifies it, the platform ranks candidate interventions by modelled return, and a dispatched order is later checked for whether it worked.
02 · How it works
- 01
Shipped five surfaces on Google Cloud Run, a command view, interventions with expandable rationale, corridor analysis, citizen reporting and outcome verification, over a 15,360-cell satellite grid.
- 02
Used difference-in-differences against control regions to verify outcomes, so the platform can state that an order worked rather than assume the drop was caused by it.
- 03
Designed an AI product whose output is an enforceable order, not a dashboard: rank interventions by modelled return, dispatch, then measure the result against control regions.
- 04
Made an untrusted crowd signal usable by gating every citizen report through a vision model before it enters the record.
03 · What it cost, and what it returned
Chose architecture around latency and unit cost, keeping PyTorch out of the request path and serving five surfaces from an embedded DuckDB file on Cloud Run.