Building an explainable AI platform that helps district administrators allocate resources and farmers make better crop decisions using Gemini, Vertex AI, BigQuery, and Google Cloud.
Climate disasters are not just weather events. They are decision problems.
When forecasts predict a strong El Niño, governments do not simply need more data. They need answers to questions like:
- Which districts will be affected first?
- Where should limited water resources be sent?
- Which crops are likely to fail?
- What should farmers sow instead?
- Why is the AI recommending this action?
Existing dashboards provide plenty of charts.
Very few provide decisions.
That became the motivation behind El Niño 2026 Decision Copilot, an AI-powered decision intelligence platform built during the Google Cloud Gen AI Academy APAC Hackathon.
The Problem
India depends heavily on the monsoon.
A severe El Niño can lead to:
- Rainfall deficits
- Reservoir depletion
- Groundwater stress
- Crop failures
- Rising food prices
- Rural employment challenges
The information already exists across dozens of government portals, weather services, satellite datasets, and agricultural reports.
The challenge is that it is scattered.
District collectors do not have time to manually combine:
- Weather forecasts
- NDVI satellite imagery
- Reservoir levels
- Mandi prices
- Contingency plans
- Drought indicators
Farmers face an even bigger challenge.
Most need a simple answer:
Given my district, should I plant the usual crop this season?
The Goal
Instead of building another dashboard, I wanted to build an AI system that reasons over multiple data sources and produces explainable recommendations.
The platform serves two audiences through the same intelligence engine.
District Administrators
They receive:
- District risk scores
- Interactive risk maps
- Reservoir outlook
- Crop stress indicators
- Resource allocation recommendations
- AI-generated explanations
Instead of simply showing that a district has high risk, the system explains why.
Farmers
Farmers interact with a conversational AI.
They can ask questions such as:
Should I grow paddy this season?
or
Which crop is safer in my district?
The assistant retrieves official contingency plans, combines them with weather forecasts, and generates localized recommendations backed by citations.
The Core Idea
Rather than building one giant AI model, I designed the platform as a decision pipeline.
Data Sources
↓
Data Ingestion
↓
BigQuery
↓
Risk Model
↓
AI Agents
↓
Explainable Decisions
Everything flows through a single decision intelligence core.
Architecture
The platform runs entirely on Google Cloud.
Data Layer
Multiple public datasets are collected continuously, including:
- Google Earth Engine
- OpenWeatherMap
- Agmarknet
- CWC reservoir reports
- CGWB groundwater data
- NRSC datasets
- ICAR-CRIDA contingency plans
- MGNREGA employment data
All datasets are standardized before being stored in BigQuery.
BigQuery as the Source of Truth
BigQuery stores district-level information including:
- Rainfall
- Weather forecasts
- Reservoir status
- Satellite vegetation indices
- Soil moisture
- Mandi prices
- Employment indicators
Every downstream component reads from BigQuery instead of directly querying external APIs.
This keeps the architecture modular, scalable, and reproducible.
Building the Risk Model
The first version uses BigQuery ML.
Instead of training a black-box neural network, I intentionally chose an interpretable approach.
The model combines features such as:
- Rainfall anomalies
- Forecast deficit
- NDVI
- Soil moisture
- Reservoir storage
- Groundwater stress
- Agricultural indicators
The output is a district-level risk score.
That score becomes the starting point for the AI agents.
Why Multiple AI Agents?
A single prompt quickly becomes difficult to maintain.
Instead, I divided responsibilities into specialized agents using Google's Agent Development Kit (ADK).
Triage Agent
Its responsibility is to answer:
Which districts require attention first?
It ranks districts and explains the factors influencing each score.
Allocation Agent
Once priorities are known, this agent decides how available resources should be distributed.
Examples include:
- Water tankers
- Emergency funds
- Drought relief supplies
Unlike pure LLM reasoning, resource allocation is deterministic so recommendations remain consistent and auditable.
Farmer Advisory Agent
This agent combines:
- District data
- Weather forecasts
- Official contingency plans
- Retrieval-Augmented Generation (RAG)
Farmers receive localized advice supported by official documents instead of hallucinated responses.
Scaling Beyond a Prototype
The original hackathon prototype focused on 23 high-risk districts.
Scaling nationwide turned out to be far more challenging than building the AI itself.
The platform now covers:
- 763 districts across India
- Weather forecasts for every district
- Nationwide satellite vegetation data
- Reservoir coverage through regional aggregation
- Hundreds of district contingency plans indexed for retrieval
Expanding reliable public datasets was significantly harder than writing prompts.
Explainability Was a Requirement
One principle guided every feature.
Never ask users to trust the AI blindly.
Every recommendation answers:
- Why?
- Which data was used?
- Which document supports the recommendation?
- Which factors increased the risk score?
For government decision-making, transparency is often more valuable than model complexity.
Technology Stack
Google Cloud
- BigQuery
- BigQuery ML
- Vertex AI
- Vertex AI Search
- Cloud Run
- Firebase Hosting
- Cloud Storage
AI
- Gemini 2.5 Flash
- Vertex AI Search
- Retrieval-Augmented Generation (RAG)
- Agent Development Kit (ADK)
Backend
- Python
- FastAPI
Frontend
- React
- Vite
- React Leaflet
Challenges
Public Data Is Not Standardized
Every source uses different schemas, update frequencies, and district identifiers.
A significant portion of the project involved cleaning, validating, and reconciling datasets before any AI processing could begin.
Explainability
Government users need confidence in every recommendation.
Every output had to include evidence rather than relying on confidence scores.
Scaling RAG
District contingency plans vary significantly in formatting and quality.
Building an automated indexing pipeline was necessary to move beyond a small proof of concept.
What I'd Improve Next
There are several areas I would continue developing:
- Retrain the risk model using nationwide datasets
- Automate ingestion using Cloud Scheduler and Cloud Functions
- Add multilingual support for Indian regional languages
- Strengthen authentication between administrator and farmer workflows
- Expand the knowledge base with additional government guidelines
What I Learned
Building AI for public-sector decision making is very different from building a chatbot.
The hardest problems were not prompt engineering.
They were:
- Data quality
- Explainability
- Deterministic reasoning
- Trustworthy recommendations
- Integrating multiple systems into one coherent workflow
Large language models become far more useful when they serve as the reasoning layer instead of the entire application.
Demo
🌐 Live Demo
https://climate-resilience-in.web.app
🎥 Demo Video
If you are building AI systems for climate resilience, agriculture, disaster management, or public-sector decision intelligence, I would love to hear your thoughts and feedback.
Happy building! 🚀
Top comments (0)