TL;DR: NASA satellites see landfill fires in Indian cities every year, but nobody tells the people living downwind. Smoulder turns those detections plus hourly wind into one answer for any address: how many days was this place downwind of a burning landfill? It runs serverless on AWS for about $3–6 a month.
The problem
There's a school in Delhi a few metres from the Bhalswa landfill. When the landfill burns, a satellite logs the fire and the detection lands in a public NASA dataset by morning.
The parents never find out. The data exists, but nobody turns it into the one question a family would ask:
"On how many days was my street downwind of a burning landfill?"
What Smoulder does
Type a place, get a number.
- Kharki, Kolkata (beside the Dhapa landfill): 34 smoke days
- Bhalswa Jahangirpuri, Delhi: 12 smoke days
Every day comes with a receipt: which satellite saw the fire, when, and which way the wind blew, hour by hour. Residents can also subscribe to an email alert when a new fire's smoke is heading their way.
Why you can trust the number
Getting a number was easy. Getting one we'd defend took three strict rules.
1. The pixel is bigger than the dump. A satellite fire pixel at Bhalswa covers about 30 ha; the whole landfill is about 28 ha. So we only count a detection if the pixel lands on or overlaps the landfill boundary. Nearby fires are shown but never counted.
2. Some "fires" are factories. At Pirana (Ahmedabad), NASA flagged most detections as a static industrial heat source. We dropped the site, and our headline went from 311 fire days to 108. We'd rather publish 108 we can defend than 311 we can't.
3. Every number says what it is. Satellite detections are labelled MEASURED; smoke days are MODELLED. The wind rule is one sentence: downwind within ±30°, at least 1 m/s, for at least an hour in the six hours after the fire.
What the data shows
- 108 fire days detected across 9 landfills in 5 cities
- 101 of them had no news report we could find
- ~12 lakh people live within 1 km of these sites, including 1.6 lakh people and 13 schools around Bhalswa alone
We checked ourselves against the news too: satellites caught 3 of 5 reported Delhi fires, and Sentinel-2 imagery confirmed about 17.6 ha of burn-like change after a major fire at Chennai's Perungudi dumpsite.
How it's built on AWS
NASA fire data · wind · Sentinel-2 (AWS Open Data)
│
EventBridge (every 30 min) → Step Functions → Lambda → S3 + DynamoDB
│ │
SNS email alerts API Gateway → Amplify web app
- EventBridge Scheduler + Step Functions run a Watch every 30 minutes: poll NASA for new fires, check the wind forecast, and alert subscribers via Amazon SNS.
- A second Step Functions Backfill rebuilds the full record on Lambda (one container image, 19 stages) and stores it in S3 and DynamoDB.
- Sentinel-2 and elevation data are read in place from the AWS Open Data registry; we never download a scene.
- Amazon Location turns addresses into points without storing them; API Gateway and Amplify serve the app.
Two design choices worth stealing:
FetchCities:
Type: Map
MaxConcurrency: 1 # NASA's rate limit, enforced by the architecture itself
- Rate limits in the architecture, not in retry loops: the city loop can't exceed NASA's quota, and every response is cached in S3.
- Alerts can't double-send: each (subscriber, fire) pair is claimed with a conditional DynamoDB write before SNS sends anything.
An AI that can't invent numbers
Amazon Bedrock drafts a short plain-language brief for each landfill, but it only gets a JSON of facts, and every number in its output is checked against that JSON. If a single number doesn't match, the draft is rejected and a template written from the same facts is shown instead. The model writes the prose; it never supplies the facts.
What we learned
- Start from a person's question, not from the dataset.
- Decide what not to count, and write the reason down.
- Label evidence honestly: measured and modelled shouldn't look the same.
- Put limits into the architecture, not into error handling.
- Wrap generative AI in checks you can show.
Data: NASA FIRMS · Open-Meteo · Copernicus Sentinel-2 and DEM (AWS Open Data) · © OpenStreetMap contributors · WorldPop
Loading article