I built a PM2.5 model for the half of India with no air monitor, and shipped it on AWS in 48 hours
AirGap predicts PM2.5 for the 368 Indian districts with no air monitor, validated on unseen cities and deployed on AWS Lambda and Amplify. What I built, what broke, and what I'd change.
I built a PM2.5 model for half of India with no air monitors, and shipped it on AWS in 48 hours
India has about 570 public PM2.5 monitors for 1.4 billion people. When I counted, 368 of the 735 districts had no monitor within 50 km. For the people there, every air-quality app shows a blank, or quietly interpolates between faraway dots without saying how wrong that is.
So for the MakeDevs Bharat Builds Tour (Track 1: Air) I built AirGap: a daily PM2.5 estimate for every district, with an error bar you can check, the safest hours to go outside, advice in English and Hindi, a 3-day forecast, and a ranked list of where the next 20 monitors should go. Live: https://main.d3cvwqcst393r5.amplifyapp.com · Code: https://github.com/akshitabiju/AirGap
Here is how it works, what I ran on AWS, and what broke.
The idea in one paragraph
You can't measure where there's no sensor, but you can learn. For each of the 571 real monitors on OpenAQ, I have daily PM2.5 readings for the last 92 days, and for the same place and day I have what Copernicus's CAMS model says about aerosol over it (satellite-assimilated PM2.5, AOD, dust) plus ERA5 weather, all via Open-Meteo with no key. A LightGBM model learns the mapping. For a district with no monitor, it reads today's satellite and weather and returns the number a monitor there would likely have shown.
Validation that answers the real question
A random train/test split would let the model see Delhi on Monday and be tested on Delhi on Tuesday. That's not the question. I used leave-cities-out cross-validation: group monitors by city, hold out whole cities, predict them cold, five folds. On cities the model never saw: mean absolute error 10.2 µg/m³ versus 16.0 for the raw CAMS baseline scored on identical rows, better in every fold. Two quantile models give a 68 % interval, and a single multiplier fitted out-of-fold scales it so the interval really covers 68 % of held-out values (it covered 48 % before calibration; an error bar you can't check is decoration).
The AWS part
AWS Lambda with a Function URL is the live piece. When a user pins any district, the browser calls
GET /?lat=&lon= . The function fetches today's CAMS and weather features for that exact coordinate, evaluates the three LightGBM model files with a pure-Python evaluator of about 79 lines (verified equal to lightgbm in tests, so the package is 2 MB instead of 100 MB), and returns the prediction and interval in about a second. No API Gateway, no containers.AWS Amplify Hosting serves the React site and its generated data files over HTTPS with a CDN. I deploy with manual zip uploads from a script; a retrained model goes live with one command.
Amazon S3 and IAM round it out: S3 for the build artefacts (with an S3 + CloudFront path scripted for accounts that have it), IAM for a least-privilege Lambda execution role. Everything is driven by the open-source AWS CLI v2 in idempotent scripts.
What broke, honestly
- CloudFront refused. New accounts can't create distributions until AWS Support verifies them. There's no self-service path, so I moved hosting to Amplify in half an hour and kept the CloudFront script for later.
- The public Lambda URL returned 403 with a correct-looking policy. Auth type NONE needs both
lambda:InvokeFunctionUrlandlambda:InvokeFunction(scoped with--invoked-via-function-url). The CLI error gives no hint. Oneadd-permissionfixed it.
- IAM onboarding. The hackathon user needed
iam:CreateRole,iam:AttachRolePolicyandiam:PassRolejust to give a Lambda a role. The AccessDenied message names the action but not the minimal policy to add.
- Not AWS, but worth sharing: Open-Meteo's archive endpoint has an hourly cap that a 1,300-point backfill hits, so cache every response and build a resume mode before the fetch, not after. And I got an OpenAQ key suspended by retrying through 429s with three workers. Respect
Retry-After. One worker.
What I'd change with another week
Hourly ground truth so the "best hours" are learned rather than borrowed from the CAMS diurnal shape; a winter retrain (monsoon PM2.5 is low and flat, so R² is modest even though the error beats the baseline by 36 %); NASA FIRMS fire detections as a feature for stubble-burning season; and a scheduled daily run on AWS instead of a laptop.
If you work on air quality in India and want the model's outputs for your district, the pipeline re-runs from open data with one command. Issues and PRs welcome.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article