Building Jaltaranga: A Universal Cloud Platform for Global Flood Evacuation Routing
How we engineered a geography agnostic serverless platform on AWS to predict street level flood depths and compute safe evacuation routes for any river basin on Earth, from Himalayan flash floods to urban deluges.
The Global Blind Spot in River Disasters
Across the globe, climate volatility is overwhelming traditional flood defense systems. In recent months, devastating torrential deluges overwhelmed rivers across Central Europe, flash flood surges originating in the Himalayan foothills of Nepal poured across international boundaries into northern India, and the Yamuna River in Delhi reached an all time high of 208.66 meters.
Whether looking at the Mississippi in the United States, the Danube in Europe, the Mekong in Southeast Asia, or the river networks of South Asia, the core systemic failure during extreme inundation is identical everywhere on Earth.
Government authorities and hydrological agencies monitor water at macroscopic checkpoints. They publish warnings like gauge heights at a regional bridge or reservoir. While that data is critical for hydrologists, it is practically useless to a family trapped in a low lying settlement or a rescue commander deploying boats on the ground. A single river stage reading does not answer the real questions that save lives: When will water enter our specific street? Which roads are completely impassable? Where is it safe to walk right now?
Worse still, everyday consumer navigation applications like Google Maps are completely flood blind. They optimize for peacetime traffic. When riverbanks breach, their algorithms actively guide fleeing residents straight into submerged culverts and underpasses, creating lethal drowning traps.
To solve this, our team built Jaltaranga for the Bharat Builds Tour 2026 Environmental Hacks hackathon. Jaltaranga is not a single city tool or an isolated prototype. It is a universal, geography agnostic decision support platform built on AWS. It translates upstream barrage discharge and river telemetry into street level water depths, computes dynamic evacuation routes around submerged roads, and broadcasts verified multilingual alerts to citizens and commanders.
Here is how I designed our machine learning forecasting pipeline and cloud architecture on AWS to be completely decoupled from geography and ready for any river basin in the world.
A Geography Agnostic Hydrological Modeling Pipeline
Traditional flood mapping relies on heavy numerical partial differential equation solvers that take six to twelve hours to run across a regional grid. When sudden upstream dam releases occur during a storm, emergency teams do not have hours to wait for compute clusters.
I formulated our river stage forecasting engine as a modular Autoregressive Ridge Lag Regression pipeline. Fluid dynamics and wave propagation obey universal physical laws whether water is flowing down the Rhine, the Bagmati in Nepal, or the Yamuna in Delhi. Downstream stage height is governed by two universal forces:
- Autoregressive momentum of local river stages across zero, one, two, three, six, twelve, twenty four, and forty eight hour lags.
- Hydrodynamic wave transit pulses travelling downriver from upstream reservoirs and barrages across twelve, twenty four, thirty six, and forty eight hour transit intervals.
Because downstream and upstream telemetry exhibits strong temporal collinearity, I applied Ridge regularization with cross validation to ensure robust generalization across different river geometries.
To evaluate our model against gold standard hydrological benchmarks used by global agencies like NOAA and USGS, we evaluated our predictions on extreme surge waves using Nash Sutcliffe Efficiency and Root Mean Square Error. Our champion model achieved:
- A Nash Sutcliffe Efficiency score of 0.942 on a six hour lead horizon, compared to 0.812 for standard persistence baselines.
- A Root Mean Square Error of just 0.18 meters at peak flood stage.
- Quantified eighty percent confidence interval coverage bands that give disaster commanders real bounds of uncertainty rather than false point precision.
Because our mathematical formulation is completely decoupled from any single river, deploying Jaltaranga to an entirely new watershed anywhere on Earth requires only two inputs: upstream discharge telemetry and downstream gauge time series.
Universal Inundation Constrained Routing on OpenStreetMap
Predicting water heights is only half the battle. The true test of disaster technology is getting people out of harm's way without leading them into drowning traps.
To make Jaltaranga deployable anywhere in the world, we built our spatial pipeline entirely on open, global geospatial datasets:
- Global digital elevation models from Copernicus and NASA SRTM rasters.
- Worldwide street networks from OpenStreetMap, which covers virtually every road and alleyway on the planet.
My teammate Saksham built our topological routing engine, which executes a time dependent Dijkstra algorithm where traversal costs increase quadratically with water depth. Any street edge where water reaches thirty centimeters, the physical limit where wading becomes hazardous for children and the elderly, is instantly assigned an infinite penalty cost.
When we evaluated our router on historical extreme surge data, the life saving impact became clear. Standard shortest path routing would have directed fifteen thousand one hundred people straight across severed transit corridors like the Ring Road. Jaltaranga automatically identified severed paths, calculated safe alternative corridors, and guided one hundred percent of residents to relief camps with verified bed availability.
Building for Global Scale on AWS
Disaster infrastructure must be serverless, multi region ready, and zero cost when idle. We built our entire backend on AWS in the ap south 1 Mumbai region using AWS Serverless Application Model, structured so it can be deployed to any AWS region globally in minutes.
1. Zero Idle Compute with AWS Lambda Graviton3
Our machine learning models, topological routing algorithms, and dispatch engines run on AWS Lambda powered by ARM64 Graviton3 processors. This gave us sub hundred millisecond execution times and instant cold starts while keeping operational costs at zero when no flood is occurring.
2. High Throughput State with DynamoDB On Demand
Live relief shelter bed counts, evacuation queue ranks, and field reports are stored across Amazon DynamoDB tables running in on demand mode. Because environmental emergencies generate sudden spikes in activity, pay per request scaling eliminated capacity planning while delivering sub millisecond database reads.
3. Multi Tier Geospatial Data Lake on Amazon S3
We store hydrodynamic flood GeoJSON polygons, elevation rasters, and synthesized voice files in Amazon S3, structured as a scalable geospatial data lake that feeds both our Lambda microservices and our web consoles.
4. Hallucination Proof Multilingual AI with Amazon Bedrock and Polly
Emergency communication cannot tolerate artificial intelligence hallucinations. If an artificial intelligence model invents a shelter location or emergency phone number during a crisis, lives are lost.
We used Amazon Bedrock with the Nova model to generate natural, localized advisory copy, but we bounded it with strict code enforced slot contracts. Bedrock writes the human sentences in any regional language, while shelter locations, walking distances, and official emergency numbers like 1077 and 112 are completely immutable and injected deterministically by code.
Once approved by an officer, Amazon Polly synthesizes the advisory into natural speech using neural voices like Kajal, ready for local community loudspeakers and mobile voice notes.
5. Real Time Alert Broadcasts with Amazon SNS
To ensure immediate notification, our dispatch approval engine triggers Amazon Simple Notification Service. The second an Incident Commander authorizes an evacuation order, SNS fans out the alert directly to citizen and responder Gmail inboxes within three seconds, delivering exact water levels, severed road warnings, and shelter directions.
6. Edge Resilience with Offline Progressive Web Apps
During extreme floods, power grids fail and cellular towers lose battery backup within hours. My teammate Arjav built our citizen mobile interface as an offline first Progressive Web App. Using Service Workers and IndexedDB, colony evacuation packs under one hundred kilobytes gzip are cached locally on the user phone, allowing turn by turn routing to function even in a total communication blackout.
Climate Adaptation as a Planetary Infrastructure Layer
Building Jaltaranga proved to our team that effective climate adaptation cannot rely on fragmented, city specific tools. Whether responding to flash floods pouring out of the Himalayas, monsoon inundations in Indian cities, or river breaches in Europe and the Americas, communities everywhere need an integrated, predictive intelligence layer.
AWS provided the robust cloud building blocks to turn raw hydrological physics, open global geospatial data, and generative artificial intelligence into a life saving platform.
Natural disasters will continue to strike our planet, but sending citizens into flooded blind spots is a preventable failure. With predictive cloud architecture, we can ensure every community on Earth receives the warning and the safe path they need to survive.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article