Do we walk to school today? One answer for one place in Delhi's air, built on AWS
On the morning after Diwali, one Delhi monitor read AQI 500 while the city's official pollution stage was II. This is how I built, on AWS, an app that gives one answer for one place, and keeps its AI from ever inventing a number.
It is 8 am on 21 October 2025, the morning after Diwali, in South Delhi. A school bag sits by the door. A parent opens a phone and finds one air-quality number for the whole city. The question is small and urgent: do we walk to school?
The parent is a sketch. The morning is real. The nearest monitor with a fresh reading, Lodhi Road, 3.8 km from a home near R K Puram, read PM2.5 at 980 µg/m³. On the US Air Quality Index (AQI) that is 500, the top of the scale. The official pollution stage for all of Delhi-NCR that morning was GRAP Stage II, set from a city-wide average of 351. GRAP is Delhi's Graded Response Action Plan.
Replayed at 8 am, The Vayu Project says: avoid going out, and wear a well-fitted N95 if you must. It names the monitor, its distance and the reading's age.
It is free and live: https://d3lmq1mze2gup3.cloudfront.net . Press Try without login to play that morning yourself.
The idea
Delhi has many air-quality numbers and none of them says what to do. A city figure cannot see one street. The nearest monitor may be kilometres away and its reading hours old.
The Vayu Project turns public air data into a decision for one named place: a home, a school, a park. Walk or stay in. Mask or not. Windows shut or open. School assembly inside or on the ground. Every number says which monitor it came from, how far away it is and how old the reading is.
I built it solo for the AWS Environmental Hacks Air track, with Claude as a coding assistant. I chose the problem and the rules, read and ran every change, built the hardware, and deployed everything to my own AWS account.
Six screens of the app: My Air, the recorded days, the map, AIR Lens, the assistant and Exposure_
What you get
You add the places you care about. Each card shows the colour of the air, the AQI, the weather, the cleanest hour, the next two days, and the monitor's name, distance and age. Alerts cover the GRAP band, smog, stubble burning, rain and wind. Around the cards sit a map of every monitor the app can read (more than 7,000 at any time), an Exposure view of a person's day, AIR Lens, which draws the air as particles in front of your camera, and an assistant in English and Hindi, by text or voice.
Three trust rules sit under every screen:
- Old readings give no advice. A reading more than six hours old is dropped, and the card shows an estimate from a regional model and says so.
- Every number is labelled Live, Estimate or Measured.
- A GRAP stage is my estimate, never an order. The card says "the air here is in the Stage X band", and the official order comes first. A school is only ever "likely" closed.
The third rule matters because places differ. On 9 November 2025, two places 4 km apart averaged 434 and 319 on India's scale over the previous 24 hours: two different GRAP bands. One city number hides that gap.
Seven real days, no waiting for a bad one
Try without login plays seven real past days through the same app, as if it were 8 am on each:
- 7 Oct 2025: rain, and the cleanest of the seven.
- 21 Oct: the morning after Diwali, AQI 500 at the nearest monitor.
- 31 Oct: the stubble-burning peak, with satellite fires upwind.
- 9 Nov and 22 Dec: two smog days.
- 23 Jan 2026: heavy rain clears the smog overnight.
- 30 Mar 2026: a windy day.
The air comes from four Delhi monitors in OpenAQ's public archive, the weather from Open-Meteo's archive and the fires from NASA FIRMS. Two written rules clean it: impossible values are dropped, and so is one monitor's hour that reads over three times the median of the others. The 86 hours dropped as lone-monitor spikes are counted in the file. The hours after 8 am are what was recorded next, so the days show the features, not forecast accuracy.
Code decides, AI only words
An air app can do harm in its assistant, so it follows one rule: fixed rules decide the advice, and the model may only word it.
Amazon Bedrock runs Amazon Nova Micro behind a Guardrail. It gets the rules' answer and the facts behind it. It may reword the answer, or answer a question the rules do not cover from those facts. Then code checks the wording:
- every number must be one of the facts;
- it may not spell numbers out or add reassurance the rules did not write;
- it may not call the air good unless the level is Good;
- masked names such as
[place]are put back, in any case.
On any doubt, the rules' own text goes out. Hindi answers never reach the model: they are wording a person wrote, with the facts slotted in. Amazon Polly, with the neural voice Kajal, reads answers aloud in both languages.
One honest gap: the number check is a set check. It proves each number came from the facts, not that it sits in the right place. That is why a reworded answer may only use the numbers in its own rule's answer.
How it runs on AWS
The app and API are one AWS CDK stack in TypeScript, in Mumbai (ap-south-1). The browser never calls an air-data provider: readings, forecasts and fires come through my API. Only place search, map tiles and Cognito sign-in are called directly. The API reads public data (OpenAQ, Open-Meteo, NASA FIRMS), applies the rules and returns one answer with its source and age.
The request path: the phone asks CloudFront, which passes /v1 to API Gateway and one Lambda, backed by DynamoDB_
Service | What it does here, and why |
|---|---|
Amazon CloudFront + Amazon S3 | One address for the app and the API, so no CORS. The bucket is private. The API says how long an answer may be kept, up to 15 minutes; an answer without that header, such as an error, is never kept. |
Amazon API Gateway (HTTP API) | A proxy, a throttle and a JWT authorizer, cheaper per request than a REST API. 25 requests a second (burst 50); 3 for the assistant and speech; 1 for forms. |
AWS Lambda | One function for every route (Node 22, Graviton, 256 MB, 10-second limit). It checks each request and answer against a schema, so a bad answer is an error, never a number. |
Amazon DynamoDB | On-demand, with TTL. The cache, saved accounts, cached speech and the daily caps. Each cap is one conditional UpdateItem : add one only if there is room, so two requests at once cannot both take the last slot. No lock. |
Amazon EventBridge Scheduler | Every 15 minutes, four reference places kept warm. Every 30, about 7,200 monitors read for the map. Retries are zero, and an alarm fires if a job stops: after 45 minutes for the first, 90 for the second. |
Amazon Bedrock with Guardrails | Nova Micro rewords; it never decides. One Converse call, temperature 0.2, 220 tokens at most, raced against a 6-second timer. |
Amazon Polly | Answers read aloud, cached by their text. |
Amazon Cognito | Optional accounts. A JWT authorizer guards the one account route, and saved data is keyed by the user id inside the verified token. |
Secrets Manager, Parameter Store, CloudWatch, SNS, X-Ray | Keys out of code; one parameter that switches the model off within a minute, with no deploy; 8 alarms; a trace of every AWS call. |
The AWS open source I used is AWS CDK and Powertools for AWS Lambda (Logger and Tracer). CDK assertion tests turn safety rules into failing builds: every Lambda on arm64, no function time limit over 2 minutes, buckets block public access, and only the account route behind a token.
Also deployed and tested: an AWS Step Functions alert loop that waits for a tap. The whole repository runs 1,285 automated tests, plus lint, types and
cdk synth , in one command.The room: a measuring unit and a cleaner
Outdoor monitors cannot see inside a room, so I built a small measuring unit: an ESP32-S3 board and a Sensirion SPS30 laser particle sensor. It sends an 18-byte packet every 2 seconds over Bluetooth, and the web app reads it with Web Bluetooth. The room's air appears on My Air beside the air outside. On 11 October the serial monitor showed my room at 18 to 22 µg/m³ PM2.5, and the live site showed the same.
Bluetooth works in Chrome on a Mac; iPhone Safari cannot do it.
The room page can show how fast the air clears when a purifier goes on, so I built a prototype cleaner: a free box, a Hicool BCF-250 centrifugal fan (₹3,300), a Smart Air pleated filter (₹1,500) and a foam pre-filter (₹200). ₹5,000 of parts.
The cleaner built in five steps: the box, the fan, the pleated filter, the foam, assembled_
A handheld vane anemometer at the outlet read about 191 cfm, roughly 325 m³ an hour.
That is airflow, which is not the same number as CADR, the rating purifiers carry. The cleaner runs on mains power and has had no electrical safety review, so please do not copy the wiring.
What went wrong, and what it taught me
My own screenshot was wrong
Writing this post, I looked at my screenshot of the morning after Diwali: AQI 500, beside "GRAP Stage II". Those do not go together. The estimate had averaged the last 24 hours the monitor had, so gaps pulled in older, cleaner hours. It also stretched one hour's PM10-to-PM2.5 ratio over the whole day.
Now it uses the last 24 hours only, needs 16 of them, caps the ratio at 2, and treats a missing PM10 as a floor. I checked it against the monitors' own PM10 on 24 station-days: 22 landed in the same stage and 2 gave no estimate because too few recent hours were there.
What I learned: a number can be right while the sentence beside it is wrong. Read your own screenshots as a stranger would.
The model never answered
My first live questions came back in the rules' plain wording, never the model's. The Guardrail's prompt-attack filter was blocking my own rewording instruction, because it sat in the user turn. With the trace off, the response said only that the guardrail had stepped in, not which filter. Moving the instruction to the system prompt fixed it.
What I learned: put instructions where the model expects them, and turn the guardrail trace on from the first call.
"Good air quality" at AQI 61
A script that drives the live site caught a model answer calling Moderate air "good", and a masked place name that was never put back. Both are now refused or restored, and each is a test. Earlier, my number check had let "[person 1]" through as the number 1.
What I learned: a number check is not a claim check. Now the wording is checked for what it claims, too.
A 403 from CloudFront, and a deploy that could not roll back
CloudFront forwarded the viewer's Host header and API Gateway answered 403. An origin request policy that forwards only the viewer address fixed it. My first deploy failed and its rollback failed too: an API stage setting named a route that did not exist yet. It had passed
cdk synth and every test.What I learned: a green
cdk synth proves the template is valid, not that CloudFormation can order it. I test the live path early, and I deploy to see the order.The sensor went quiet when Bluetooth connected
The serial log showed good readings, then nothing the moment the app connected. The sensor had stopped measuring, most likely from a power dip. The firmware now restarts it on connect and after three failed reads.
What I learned: hardware fails on power, not on logic. I now read the serial log first, and the firmware assumes the sensor can drop out.
Numbers
- Cost: about $4.50 a month idle, worked out from AWS list prices, not read from a bill. An assistant question costs $0.0008, and 98% of that is the guardrail, so the caps sit on the assistant. Everything at its cap comes to about $46 a month.
- Speed: over 24 hours the API's Lambda took a median 1.5 s and 5.2 s at p99 (a place it has not seen is fetched live from OpenAQ). I saw 1.6 to 7.3 s end to end on a fresh lookup, including one cold start. A repeat comes back from CloudFront in about 0.15 s.
What is next
I want to take it into a school for a term and see what changes.
What I learned
About AWS: Bedrock in Mumbai needs an inference profile and a two-ARN policy, and its Guardrail reads the user turn for attacks, so instructions go in the system prompt. A DynamoDB conditional write is a race-free quota with no lock. CloudFront cache policies let the API set each answer's lifetime, and the origin request policy decides what API Gateway sees. CDK assertion tests turn safety rules into failing builds, but only a real deploy shows what CloudFormation can order and roll back. EventBridge retries are a setting, a silent schedule needs its own alarm, a Systems Manager parameter makes a fail-closed kill switch, and the Price List API showed me where the money goes, so the caps went there.
Beyond AWS: keep the model's job small, and read your own screenshots as a stranger would. A number can be right and the sentence beside it wrong.
Try it: https://d3lmq1mze2gup3.cloudfront.net
Seven recorded days: https://d3lmq1mze2gup3.cloudfront.net/#/week
Code: https://github.com/hrsh-kr/the-vayu-project
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article