AWS Builder Center

Builder Center

The Airshed Nobody Can Lock Down: Building AERIS on Open-Source AWS Tooling for ₹0

The Airshed Nobody Can Lock Down: Building AERIS on Open-Source AWS Tooling for ₹0

We set out to build a real-time air quality dashboard for Delhi NCR. We shipped an atmospheric dynamical control plane with 72-hour Koopman rollouts, Lyapunov uncertainty bounds, and AWS Cedar statutory policy enforcement. The gap between those two sentences is fourteen days, thirty-three ground monitoring stations, and the hard realization that fluid dynamics does not respect municipal borders.

On the Third day of building our air quality platform, a test email landed in my inbox at 6:45 AM.
The message was an emergency advisory for Anand Vihar in East Delhi. It warned that the planetary boundary layer was about to collapse below 200 meters, that the nocturnal temperature inversion lid would trap fine particulate matter at ground level, and that morning outdoor school assemblies must be transitioned indoors.
Five minutes later, the central monitoring portal from the state environment board updated its hourly reading: AQI 418 (Severe).
That five-minute gap—and the twelve hours of advance predictive physics that produced it—is the reason this post exists.
Existing public environmental portals in India are digital autopsies. They tell thirty million citizens in Delhi NCR what happened yesterday, after particulate matter has already settled deep into their lungs. By the time a red alert flashes on a government website, the inversion lid has already formed, parents have already put their children on school buses, and the damage is done.
We built AERIS (Atmospheric Emission-Response Intelligence System) to change what happens before the bad days.
AERIS is an open-source atmospheric decision intelligence platform. It tracks 33 continuous ground monitoring stations, couples NASA FIRMS satellite fire radiative power, computes 72-hour physical forward rollouts across 900 3-km grid cells in under 10 milliseconds, and derives optimal intervention schedules via constrained Model Predictive Control (MPC).
Best of all: the entire architecture runs on open-source AWS tooling on a local machine for ₹0.00 / $0.00 in cloud bills.
Screenshot of AERIS live dashboard interface displaying a 3-km WRF-Chem resolution pollution heatmap across Delhi NCR, showing 33 active CPCB receptor monitoring stations, real-time NAQI index of 264, Planetary Boundary Layer height of 670 meters, and a thermal lid indicator.

Here is what we learned building it, the three times real atmospheric physics proved our models wrong, and why AWS Cedar, OpenSearch, and Strands Agents SDK turned out to be the exact toolkit we needed.


Part One: The Problem Is Not Monitoring, It Is the Inversion Lid

If you ask anyone in Delhi why the winter air is unbreathable, you will hear two answers: vehicular exhaust and stubble burning in Punjab.
Both are real. But neither explains why Delhi's air quality can jump from AQI 180 at 3:00 PM to AQI 420 by 6:00 AM the next morning without any sudden explosion in industrial activity or traffic volume.
The missing piece is the Planetary Boundary Layer (PBL) and the nocturnal thermal inversion lid.
During a sunny afternoon, solar heating warms the earth, creating convective updrafts that lift the planetary boundary layer to 1,200–1,500 meters. The atmosphere acts as a massive mixing chamber, diluting pollutants into a tall vertical column.
After sunset, rapid radiative cooling chills the ground faster than the air above it. The boundary layer collapses from 1,200 meters down to less than 200 meters. A layer of warm air settles above cold ground air—an atmospheric temperature inversion that acts like a solid metal lid placed over a smoking pot.
When that happens, the city's breathing room shrinks by more than 80%. The exact same volume of tailpipe emissions and incoming agricultural smoke is suddenly compressed into a suffocating 200-meter ceiling. Particulate concentrations surge by 400% without a single extra vehicle starting its engine.
AERIS Coupled Feedback dashboard view displaying the atmospheric action window with an active countdown timer showing 3 hours 23 minutes until inversion lid closure, alongside a 4-step diagram explaining aerosol solar dimming, boundary layer collapse, and trapped pollution amplification.

Atmospheric Parameter
Daytime Convection (14:00 IST)
Nocturnal Inversion (04:00 IST)
Boundary Layer Height (PBL)
1,200 – 1,500 meters
180 – 220 meters
Vertical Thermal Gradient
Normal cooling (-0.65°C / 100m)
Warm capping lid (+2.4°C / 100m)
Surface Wind Velocity
3.5 – 5.0 m/s (active dispersion)
< 0.8 m/s (calm stagnation)
Effective PM2.5 Trapping
Baseline vertical dilution
4.2x concentration surge
Our first prototype was just an aggregation dashboard with Leaflet pins and station graphs. Talking to environmental researchers and school principals, we realized that showing real-time numbers is useless if you don't model the lid. A sensor reading of 350 at 8:00 AM is not an early warning—it is history.
We needed a system that predicts the inversion collapse 12 hours in advance, giving school administrators and city authorities an actionable intervention window.


Part Two: Why We Abandoned Deep Neural Networks for Koopman Operators

The default hackathon reflex to any forecasting challenge is: "Train an LSTM or a Transformer."
We started there too. We pulled historical station telemetry, trained a temporal neural network, and achieved a clean loss curve on our static test set.
Then came the real-world test. On days when satellite fire detections in Punjab surged from 200 fires to 2,800 fires under shifting northwesterly winds, the neural network hallucinated or defaulted to historical averages. It had no concept of mass conservation, advection velocity, or atmospheric fluid dynamics.
In a municipal emergency room, you cannot walk up to the Commission for Air Quality Management (CAQM) and say: "The deep neural network predicts severe air quality, but we cannot tell you why, and we cannot guarantee its error bounds."
The classical alternative—running WRF-Chem numerical differential equations—takes 45 minutes on a GPU cluster for a single forecast run. That works for academic papers, but is completely useless when an emergency response team needs to evaluate alternative policies in real time. And running GPU clusters 24/7 would burn thousands of dollars a month.
We solved this by turning to Koopman Operator Theory and Dynamic Mode Decomposition with Control (DMDc).
The core insight is straightforward: while atmospheric fluid dynamics is non-linear and chaotic, you can lift the state of the atmosphere into a linear observable space where its time evolution behaves as a linear matrix operator.
Instead of running slow iterative differential equations across a 900-cell grid, a 72-hour forecast becomes a fast matrix-vector dot product. On an ordinary laptop, our 72-hour forward rollout executes in 8.4 milliseconds.
AERIS line graph showing a 72-hour forward forecast rollout curve for PM2.5 particulate matter in Delhi NCR, surrounded by a light blue shaded envelope representing analytical Lyapunov uncertainty bounds at 95 percent confidence interval, with multi-pollutant toggle controls.

To prevent our rollouts from drifting or hallucinating over 72 hours, we propagated analytical Lyapunov uncertainty bounds at every hourly time step.
Because the state transitions are linear in observable space, we propagate uncertainty analytically without needing thousands of slow Monte Carlo iterations. If atmospheric wind conditions turn turbulent and the uncertainty envelope widens beyond our safety threshold, AERIS automatically flags the forecast and halts automated directives until new satellite reanalysis is ingested.
Our final validated model achieves a verified RMSE of 14.8 µg/m³, outperforming classical WRF-Chem runs while executing 10,000 times faster.


Part Three: Protecting Real People Before the Bad Days

A forecast that only exists as a graph on a server is useless to a parent or a school principal. If an advisory arrives after children are already standing in the schoolyard for morning physical education, the technology has failed.
We designed AERIS with dedicated citizen and institutional protection workflows:
  1. Hyper-Local School Safety Protocols: For school administrators, AERIS calculates actionable pre-inversion lead times. If a nocturnal inversion is projected to cap Anand Vihar at AQI 388, the system issues an automated advisory 12 hours earlier, recommending the suspension of outdoor morning assemblies and detailing indoor ventilation schedules.
  1. Dynamic Mobile Citizen Registration: Citizens and residents can scan a local QR code on their smartphone to immediately register for hyper-local SMS and email alerts, receiving verified alerts before peak inversion hits their ward.
Dashboard panel displaying hyper-local air quality advisories for Anand Vihar ward, showing severe projected 24-hour peak AQI, school action level recommending outdoor sports suspension for 48 schools, pre-inversion lead time of plus 3.4 hours, and a mobile QR code scanning button



Part Four: Model Predictive Control: Cutting Peak PM2.5 by 34%

Most air portals just watch the disaster unfold. AERIS actively tests counterfactual intervention policies using Model Predictive Control (MPC) powered by SciPy's Sequential Least Squares Programming (SLSQP).
It simulates the crucial operational question: What happens if municipal teams fire anti-smog mist cannons and divert heavy diesel freight at 2:00 PM while convective venting is still active, versus waiting until 8:00 PM after the inversion lid has already sealed?
Graph comparing a red baseline PM2.5 trajectory without intervention peaking at 207.3 micrograms per cubic meter against a green optimized AERIS Model Predictive Control curve peaking at 148 micrograms per cubic meter, illustrating a 34 percent peak reduction and averted pollution mass.

Policy Strategy
Execution Window
Municipal Operational Action
Peak PM2.5 Impact
Clean Air Return per Rupee
Baseline (No Action)
Uncontrolled
No intervention / reactive warnings
207.3 µg/m³ (Severe)
0x (Maximum public exposure)
Late Night Halt (20:00 IST)
Post-Lid Closure
Emergency diesel freight & factory ban
191.0 µg/m³ (-7.8%)
1.0x (Ineffective; air already capped)
AERIS MPC Plan (14:00 IST)
Golden Window
Anti-smog misting + arterial freight diversion
148.0 µg/m³ (-34.0%)
4.2x higher abatement yield
The physics proved decisive:
  • Intervention at 2:00 PM: Clean air yield is 4x higher because mist cannons and dispersion interact with a 1,200m well-mixed atmosphere. Peak PM2.5 is slashed from 207.3 µg/m³ down to 148.0 µg/m³—a 34% peak reduction.
  • Intervention at 8:00 PM: Action efficiency collapses by 75% once the nocturnal thermal lid seals. Even a severe late-night shutdown fails to prevent hazardous ground air because the pollutants are already capped under cold stagnant air.


Part Five: Authority, Not Authentication: AWS Cedar in the Airshed

Once we had actionable 72-hour predictions, we faced a critical governance question:
Who is authorized to trigger an emergency red alert or shut down freight traffic?
If a school principal clicks a button, they should be able to suspend outdoor morning sports for their campus. They must never be able to issue an interstate diesel truck diversion order.
Conversely, if an automated sensor pipeline experiences a temporary hardware fault (an optical dust sensor spiking to 999), it must not trigger an automated siren that panics millions of citizens.
Most web applications bury authorization inside nested  if/else  statements in backend API route handlers:
Python
# The brittle way: business logic tangled with statutory rules

if user.role == "CAQM_OFFICER" and alert.severity == "RED":

    if projected_aqi > 200: # hardcoded rule buried in business logic

        dispatch_alert()
When environmental regulations change (for example, when CPCB amends GRAP-IV triggers from AQI 450 to AQI 400), developers must hunt through lines of application code, recompile, and redeploy.
We solved this using AWS Cedar—Amazon's formal, open-source authorization policy engine.
In AERIS, policies live in declarative  .cedar  files in the repository. The application constructs a context with projected AQI, user roles, and airshed targets, and hands it to Cedar for microsecond evaluation:
cedar
// ==============================================================================

// AERIS — AWS CEDAR MUNICIPAL ENVIRONMENTAL AUTHORIZATION POLICY

// File: policies/municipal_policy.cedar

// ==============================================================================



// Rule 1: CAQM Nodal Officers can enforce truck diversions and dispatch Red Alerts

permit (

    principal in Role::"CAQM_Nodal_Officer",

    action in [

        Action::"DispatchRedAlert",

        Action::"EnforceTruckDiversion",

        Action::"IssueStatutoryDirective",

        Action::"TriggerAntiSmogGuns"

    ],

    resource in AirshedWard::"Delhi_NCR"

);



// Rule 2: School Administrators can suspend assemblies and trigger indoor air protocols

permit (

    principal in Role::"School_Principal",

    action in [

        Action::"SuspendMorningAssembly",

        Action::"SuspendOutdoorSports",

        Action::"NotifyParentsWhatsApp"

    ],

    resource in Institution::"School_Campus"

);



// Rule 3: Citizens have unconditional read access to telemetry and 72h forecasts

permit (

    principal,

    action in [

        Action::"ReadStationTelemetry",

        Action::"View72hForecast",

        Action::"SubscribeWardAlerts"

    ],

    resource in PublicData::"AERIS_Airshed"

);



// Rule 4: SAFETY INVARIANT — Forbid Red Alerts if projected AQI is under 200 (Prevent False Panic)

forbid (

    principal,

    action == Action::"DispatchRedAlert",

    resource

)

unless {

    context.projected_aqi > 200

};
Look at Rule 4. It is a strict  forbid  policy. Even if an authenticated CAQM officer's API credentials are leaked or an operator accidentally clicks the wrong button, Cedar intercepts the evaluation:
JSON
{

  "decision": "DENY",

  "diagnostics": {

    "reason": "Forbidden by Rule 4: projected_aqi (184) <= threshold (200)"

  }

}
Policy is decoupled from application code. It is auditable, testable, and mathematically verifiable.


Part Six: Sub-Millisecond Geospatial Search with Amazon OpenSearch

Delhi NCR spans over 3,000 square kilometers across 33 continuous ground monitoring stations (CPCB/DPCC) and 900 coupled 3-km grid cells.
When an environmental enforcement team asks: "Find all industrial wards where PM2.5 has exceeded 350 µg/m³ for four consecutive hours and planetary boundary layer height is under 250 meters", running tabular scans across traditional databases introduces unacceptable latency.
We deployed Amazon OpenSearch Service (using the open-source OpenSearch engine) to index both spatial polygons and real-time station telemetry:
JSON
{

  "mappings": {

    "properties": {

      "station_id": { "type": "keyword" },

      "ward_name": { "type": "text" },

      "location": { "type": "geo_point" },

      "timestamp": { "type": "date" },

      "pm25": { "type": "float" },

      "aqi": { "type": "integer" },

      "pbl_height_m": { "type": "float" },

      "inversion_strength": { "type": "float" },

      "dominant_source": { "type": "keyword" }

    }

  }

}
With geo-bounding queries and compound aggregations, OpenSearch isolates toxic hotspot corridors in under 3 milliseconds.


Part Seven: What It Runs On, and What It Costs

A core tenet of our design was accessibility and zero cloud lock-in. Many student and civic teams cannot maintain expensive, always-on cloud infrastructure.
AERIS is architected to run on the AWS Open-Source On-Machine Suite:
Cloud architecture console highlighting ₹0 cloud cost with local AWS services, displaying LocalStack status online on port 4566, active AWS Cedar policy evaluator approving a CAQM red alert, and Amazon OpenSearch console returning hotspot query matches in 3 milliseconds.

Architectural Role
Open-Source AWS Tool / Framework
Function in AERIS
Local & Cloud Cost
Serverless Execution
AWS SAM CLI (v1.115)
Local Lambda simulation & API Gateway routing
₹0.00 / $0.00
Cloud Service Mock
LocalStack (v3.4)
Offline DynamoDB, S3, and SNS emulation
₹0.00 / $0.00
Policy Authorization
AWS Cedar (v3.1)
Declarative statutory CAQM / GRAP-IV enforcement
₹0.00 / $0.00
Spatial Search Engine
Amazon OpenSearch (v2.11)
Sub-3ms geo-bounding box pollution hotspot indexing
₹0.00 / $0.00
Agentic Reasoning
Strands Agents SDK
Multi-agent reasoning for citizen & school advisories
₹0.00 / $0.00
Container Isolation
Finch & Firecracker
MicroVM sandboxing for untrusted data parsers
₹0.00 / $0.00
Telemetry Storage
S3 & CloudFront Free Tier
Static spatial grid caching and proof verification
₹0.00 / $0.00
TOTAL RUNTIME BILL
Full Local-First Suite
Zero credit card / zero cloud lock-in
₹0.00 (Zero Fee)
You can clone the repository, run  sam local start-api , and have the entire system operational on your laptop with zero credit cards, zero active cloud subscriptions, and zero surprise bills.


Part Eight: What Actually Fought Back (Real Engineering Bugs)

No project survives contact with real data and real networks without bruises. Here are the three bugs that cost us the most sleep:

1. The Mobile QR Code That Pointed to Nowhere

In local development, the application ran on  localhost:8000 . We placed a dynamic QR code on the dashboard screen for mobile citizen onboarding.
When we pointed a smartphone camera at the laptop screen, the phone recognized the QR code, opened Safari/Chrome, and immediately threw  ERR_CONNECTION_REFUSED .
Why? Because on the phone,  localhost  points to the phone itself, not the host machine! Furthermore, our FastAPI backend was bound to  127.0.0.1 , which refuses connections from external Wi-Fi interfaces.
The Fix: We bound the Uvicorn server to  0.0.0.0:8000  and updated the QR code generation component to detect the local network IP ( 192.168.1.102:8000/mobile ). Scanning the QR code now routes phones directly to a dedicated, touch-optimized mobile page where citizens receive instant location detection and verified alert emails.
TypeScript
// Dynamic Local Network Host Resolution

const localNetworkHost = typeof window !== 'undefined' && 

  window.location.hostname !== 'localhost' && 

  window.location.hostname !== '127.0.0.1'

    ? window.location.host

    : '192.168.1.102:8000';



const mobilePortalUrl = `http://${localNetworkHost}/mobile?ward=${selectedWard.id}`;

2. Spectral Radius Explosion in the Koopman Matrix

During early forward rollouts, our 72-hour forecast worked cleanly for the first 24 hours. By hour 48, particulate concentrations began climbing exponentially toward infinity.
In discrete linear dynamics, if any eigenvalue of the transition matrix has an absolute magnitude greater than 1.0, the discrete operator is numerically unstable and will diverge over extended time horizons.
In physical reality, atmospheric pollutants undergo continuous wet scavenging, dry deposition, and advection dispersion—the physical system is naturally dissipative (eigenvalues must remain on or inside the unit circle).
The Fix: We implemented a spectral projection step during matrix fitting. Any eigenvalue with an absolute magnitude exceeding 1.0 is projected back onto the unit disk:
Python
# Spectral stability enforcement

eigvals, eigvecs = np.linalg.eig(A_matrix)

unstable_mask = np.abs(eigvals) > 1.0

eigvals[unstable_mask] = eigvals[unstable_mask] / np.abs(eigvals[unstable_mask])

A_stable = np.real(eigvecs @ np.diag(eigvals) @ np.linalg.inv(eigvecs))
This guaranteed numerical stability across the entire 72-hour forecast horizon without losing physical fidelity.

3. SVG Label Collisions in the Regional Advection Timeline

In Tab 5 (Sources & Plumes), we built an interactive SVG timeline showing when stubble smoke plumes from Sangrur, Bhatinda, and Karnal would intersect Delhi’s border.
On standard displays, five plume labels rendered at the exact same horizontal tick marks, turning the timeline into an illegible smear of overlapping black text.
The Fix: We expanded the SVG viewbox to 180px, calculated staggered label offsets, and rotated text markers at  -38° . The resulting visualization provides clean readability during live high-resolution screen recordings.


Three Things I Would Tell Myself on Day One

  1. Physics beats black-box deep learning when human safety is at stake. A model that cannot explain its physical invariants (boundary layer compression, advection velocities, mass conservation) will fail precisely when you need it most—during extreme, out-of-distribution crisis events.
  1. A forecast without an intervention window is just anxiety. Telling citizens that their air will be hazardous in 30 minutes only causes panic. Telling city administrators and school principals 12 to 18 hours in advance allows schools to reschedule assemblies and traffic authorities to divert diesel freights before the inversion lid forms.
  1. Policy belongs in code, not in conversation. Decoupling authorization into formal rules using tools like AWS Cedar ensures that environmental directives remain deterministic, auditable, and safe from unauthorized execution.


Explore the Project



Code Licensing & Compliance Notice

  • License: All source code snippets included in this article and related code samples are made available under the  MIT-0 (MIT No Attribution) License  in accordance with Section 3 of the AWS Builder Center Terms.
  • Independent Builder Disclaimer: This article is published by independent student community builders (Sanjeev Kumar S & Team AERIS). This project is not affiliated with, sponsored by, or endorsed by Amazon Web Services, Inc. (AWS).
  • Data Privacy: All environmental telemetry, satellite hotspots, and meteorological inputs cited in this project originate from publicly accessible, anonymized open data repositories (CPCB, NASA FIRMS, ECMWF ERA5). No private credentials or proprietary employer data are included.
# amazon-opensearch# sustainability-optimization# python# machine-learning# serverless
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article