ResilientFlow AI: Climate-Aware Infrastructure Copilot

An agentic AI system that combines reasoning, multimodal data, and optimization to support climate-resilient infrastructure investment decisions across regions.

Comment

ResilientFlow AI: Climate-Aware Infrastructure Copilot

Inspiration

Climate change is already disrupting infrastructure systems worldwide—from the Mekong Delta to the US Gulf Coast. Floods and extreme weather events increasingly threaten transportation networks, supply chains, and communities.

Current planning tools are often:

  • Static and scenario-limited
  • Hard to use for non-experts
  • Not transferable across regions

What if decision-makers could ask questions in natural language and instantly receive optimized, explainable infrastructure strategies?

What This Project Does

ResilientFlow AI is an agentic AI-powered decision platform that helps:

  • Evaluate infrastructure investments under climate risks
  • Optimize multi-period, multi-modal transport networks
  • Compare planning scenarios
  • Interact using natural language

Example query:

“With a budget of $50M, which projects reduce flood disruption the most?”

System outputs:

  • Optimal investment portfolio
  • Quantified impact (cost, disruption reduction)
  • Evidence-based explanation
    ## How It Works

Agentic Architecture (Amazon Nova)

  • Orchestrator Agent (Nova 2 Lite)
    → Parses user intent and coordinates workflow

  • GeoAdapter Agent
    → Standardizes datasets across regions

  • Knowledge Agent (RAG)
    → Uses Nova Multimodal Embeddings to retrieve from:

    • PDFs
    • Reports
    • Maps

  • Optimization Agent
    → Solves network design problem

  • Explanation Agent (Nova 2 Lite)
    → Generates human-readable insights

Optimization Model

We model the system as a multi-period infrastructure optimization problem:

$$ \min \sum_{t} \left( C_{transport}^{(t)} + C_{investment}^{(t)} + C_{disruption}^{(t)} \right) $$

Subject to:

  • Flow conservation
  • Capacity constraints
  • Budget constraints

Climate impact is modeled as:

$$ capacity_{ij}^{(t)} = base_{ij} \cdot (1 - flood_impact_{ij}^{(t)}) $$

Geography-Agnostic Design

Supports multiple regions:

  • Mekong Delta (demo)
  • Netherlands
  • Bangladesh
  • US Gulf Coast
  • EU river corridors

Users can:

  • Upload custom datasets (CSV, GeoJSON, PDF)
  • Select predefined regions
  • Automatically standardize data

Tech Stack

  • Amazon Nova (Bedrock)

    • Nova 2 Lite (reasoning)
    • Nova Multimodal Embeddings (RAG)

  • Optimization

    • Pyomo + HiGHS / Gurobi

  • Vector Search

    • FAISS

  • Frontend

    • Streamlit + Map visualization

Challenges

1. AI + Optimization Integration
LLMs cannot solve optimization directly

→ Solution: separate reasoning (Nova) and computation (solver)

2. Data Heterogeneity
Different regions use different schemas

→ Solution: GeoAdapter standardization

3. Explainability
Optimization outputs are hard to interpret

→ Solution: Nova-generated explanations

What We Learned

  • Agentic AI works best with tools
  • Decision intelligence > text generation
  • Multimodal data is critical in climate systems
  • Clean architecture is key

Future Work

  • Real-time climate data integration
  • Voice interaction (Nova Sonic)
  • GIS / digital twin integration
  • Multi-user collaboration
  • Global deployment

Final Thought

ResilientFlow AI transforms infrastructure planning into an interactive, AI-powered decision-making system for a climate-resilient future.

Built With

Share this project:

Updates