SkyGraphOps

AI-driven airline disruption management platform: forecasts delay propagation across an airport network with a spatio-temporal graph neural network, generates recovery plans with an MILP solver, and uses an LLM agent to turn both into a plain-English operations briefing.

Live demo: skygraphops.vercel.app

What it does

When a disruption hits (weather, aircraft-on-ground, crew sickness), an airline operations center needs to answer three questions fast: how bad will this get, what are my options, and what do I do right now. This project builds an end-to-end pipeline for that:

1. Forecast — a heterogeneous graph neural network predicts per-airport delay (with uncertainty bounds) from live network state.
2. Optimize — a MILP recovery solver generates multiple EASA-compliant recovery plans (reroute/delay/cancel), warm-started by a DRL policy.
3. Decide — an LLM agent (Claude) orchestrates the forecast and solver as tools, reasons over the tradeoffs, and produces a short briefing a duty manager can act on — with every step logged for auditability.

Architecture

flowchart TD
    A[Raw flight + weather data] --> B[Graph Builder]
    B --> C[Heterogeneous Graph\nairport / flight / crew / aircraft]
    C --> D[ST-GNN Forecaster]
    D --> E[Delay predictions: q10 / q50 / q90]

    F[Scenario Generator\nSNOWSTORM, AOG, CREW_SICKNESS] --> H
    E --> H[LLM Agent Orchestrator]
    G[MILP Recovery Solver + DRL warm-start] --> H
    H --> I[FastAPI state machine\npredict / scenarios / apply / rollback]
    F --> I
    E --> I
    G --> I
    I --> J[React frontend]

Loading

Layer	Path	Role
Graph	graph/	Builds the heterogeneous graph (airport/flight/crew/aircraft) from raw flight + weather data
Forecasting	forecaster/	ST-GNN model, training, evaluation, inference
Scenarios	scenarios/	Disruption scenario definitions (weather, AOG, crew sickness)
Recovery	solver/	CP-SAT / MILP recovery optimizer with DRL warm-start
Agent	agent/	Claude-based orchestrator, tool loop, XAI audit logging
API	api/	FastAPI state machine wiring forecast → solver → agent
Frontend	frontend/	React + TypeScript ops dashboard

Model performance

The ST-GNN forecaster is evaluated against ARIMA and XGBoost baselines on a fair, system-level aggregated task:

Model	Task	MAE (min)	RMSE (min)
ARIMA	System-level daily avg	5.88	6.78
XGBoost	System-level daily avg	4.31	5.32
ST-GNN	Per-airport (harder task)	6.25	9.34
ST-GNN	System-level daily avg	3.18	—

ST-GNN beats ARIMA by 46% and XGBoost by 26% on MAE once aggregated to the same task the baselines solve. Quantile heads (q10/q50/q90) additionally give risk-aware interval forecasts that feed directly into recovery planning.

Tech stack

• ML: PyTorch, PyTorch Geometric (heterogeneous GNN, quantile regression), scikit-learn baselines
• Optimization: Google OR-Tools (CP-SAT / MILP), DRL warm-start
• Agent: Anthropic Claude with a custom tool-use loop, MCP servers for forecaster/solver
• Backend: FastAPI, SQLite
• Frontend: React, TypeScript, Vite, Tailwind, Zustand, Recharts, Leaflet
• Deploy: Railway (API), Vercel (frontend)

Running locally

# Backend
pip install -r requirements.txt -r requirements-api.txt
uvicorn api.main:app --reload --port 8000

# Frontend
cd frontend
npm install
npm run dev

The API expects an ANTHROPIC_API_KEY in a .env file at the project root for the agent orchestrator to run.