EarthTwin

An AI-powered planning engine that simulates real-world environmental impact before construction begins.

Comment

🌍 Earth Twin — Plan Before You Build (sustainability track)

💡 Inspiration

Modern infrastructure projects often fail not because of bad intent, but because of late discovery. Environmental impact pollution, biodiversity loss, regulatory risks is usually identified after millions have already been spent.

Environmental impact assessments (EIAs) can take months to years, cost hundreds of thousands of dollars, and still come too late to prevent damage.

Our inspiration is grounded in real data. We analysed six years of EPA AQS + CASTNET ozone measurements (374,000+ rows) and found ozone rose nearly 20% between 2020 and 2025 a consistent, worsening trend directly linked to unplanned construction and industrial activity. The full data analysis behind Earth Twin is open source:

📊 HackAugie 2026 Data Challenge — Earth Twin Ozone Analysis Six years of real ozone data proving that construction without planning makes air quality measurably worse every year.

We asked a simple but powerful question:

What if you could simulate the environmental impact of a project instantly — before building anything?

That idea became Earth Twin.

🧠 What We Built

Earth Twin is an AI-powered environmental planning system that allows users to:

  • Describe any infrastructure project in plain English
  • Simulate its environmental impact instantly
  • Compare "with planning" vs "without planning" scenarios
  • Generate a professional planning report

Instead of replacing EIAs, we eliminate the blind guesswork before them.

⚙️ How We Built It

Our system combines three core components:

1. 🔎 RAG + Gemini (Structured Understanding)

We built a retrieval-augmented generation (RAG) pipeline that:

  • Matches user prompts to validated infrastructure templates
  • Injects domain-specific constraints (capacity, environmental rules, etc.)
  • Uses Gemini to extract structured JSON (not free text)

This ensures:

Output = User Input + Domain Constraints

2. 🧮 Deterministic Simulation Engine

Unlike typical AI systems, our environmental calculations are not AI-generated.

We model impact using controlled formulas such as:

Ozone_future  = Ozone_baseline × (1 + Δ_construction)
Planned Impact = Baseline      × (1 − Δ_optimization)

Example from our dataset:

Scenario Ozone Change
Without planning +8% increase
With Earth Twin −5% optimized reduction

This creates trustworthy, explainable results.

3. 🤖 Gemini as Planner + Analyst

We use Gemini in three distinct roles:

Role Responsibility
Interpreter Extract structured plans from natural language
Planner Decide actions based on real environmental metrics
Analyst Generate grounded reports from simulation outputs

This separation prevents hallucination and keeps AI strictly grounded in reality.

📊 What We Learned

  • AI alone is not enough — combining AI with deterministic systems creates trust
  • RAG is powerful when used for constraint injection, not just retrieval
  • The hardest problem is not generation — it's structuring inputs and outputs correctly
  • Real-world impact comes from decision support, not just predictions

⚠️ Challenges We Faced

1. Avoiding Hallucination

We solved this by:

  • Restricting Gemini to structured outputs
  • Feeding it only validated context
  • Separating simulation from AI

2. Data Realism

Environmental modeling requires real baselines, not fake data. We integrated:

3. Bridging AI + Systems

The biggest challenge was designing a system where:

  • AI makes decisions
  • But math validates them

This hybrid architecture was the core breakthrough.

🚀 Impact

Earth Twin transforms:

Before After
2–5 year environmental reviews Seconds
Expert-only tools Plain English accessibility

We are not replacing regulation — we are making better decisions before regulation is needed.

🔮 Future Scope

  • Fully mobile-optimized platform
  • Expanded infrastructure types (transport, energy, urban planning)

Earth Twin shows: *"What will happen — before you build."*

Built With

  • cesium-ion
  • fastapi
  • google-gemini-api-(gemini-2.5-flash)
  • google-genai-sdk
  • nominatim-(openstreetmap-geocoding)
  • open-meteo-api
  • postgresql
  • pydantic
  • python
  • rag-(custom)
  • react
  • render
  • supabase
  • vercel
  • vite
  • world-bank-api
Share this project:

Updates