AeroTwin

Inspiration

In airline operations, an Operations Control Center (OCC) dispatcher must sift through scattered documents—Flight Plans, Weather Briefings, Crew Schedules, Maintenance Logs—to assess a single flight. This process typically takes 15–20 minutes per flight and is prone to human error due to fragmented information. We wanted to build a system that enables AI to "understand the entire flight" rather than just "reading individual documents."

What it does

AeroTwin AI is an AI Operating System for Airline OCCs, comprising four core components:

Aviation Document Intelligence: Automatically extracts structured data from operational PDFs (Flight Plans, Load Sheets, Crew Rosters).
Flight Graph Engine: Organizes all flight-related data into a unified graph (aircraft, crew, airports, weather, maintenance status...).
Operational Risk Engine: Calculates real-time risk scores based on business rules (duty time limits, weather thresholds, maintenance status, gate conflicts).
Operations Copilot: An AI assistant that provides actionable operational recommendations based on the full context of the flight.

How we built it

Frontend: Next.js, Tailwind CSS
Backend: Node.js, Hono
Database: PostgreSQL (using nodes + edges tables for graph representation)
AI Layer: OpenAI API (OCR, information extraction, reasoning)
Storage: Cloudflare R2
Deployment: Cloudflare Workers (backend API), Vercel (frontend)

Challenges we ran into

Extracting accurate data from non-standardized aviation PDFs → Solution: Prompt engineering + schema validation layer.
Designing a Flight Graph flexible enough to represent complex operational relationships → Solution: Hybrid relational-graph model on PostgreSQL.
Balancing false positives/negatives in the Risk Engine with noisy data → Solution: Weighted scoring + confidence thresholds.
Synchronizing the pipeline (Document Intelligence → Graph → Risk Engine → Copilot) within 72 hours → Solution: Event-driven architecture with Redis Streams.

Accomplishments that we're proud of

Built a complete end-to-end pipeline from raw PDF → structured JSON → risk score → actionable recommendation in 72 hours.
Operations Copilot can answer real-world queries like "Should flight VJ123 depart on time?" with specific, cited reasons.
Risk Engine classifies risks into 4 levels (LOW / MEDIUM / HIGH / CRITICAL) with transparent, explainable scoring.
Achieved zero manual data entry for the demo flow.

What we learned

Graph models are significantly superior to relational databases for representing multi-entity airline operational data.
Gemini API handles specialized aviation PDFs better than expected when combined with structured output schemas and few-shot prompting.
Defining a strict scope from day one was the critical success factor for completing the MVP within 72 hours.

What's next for HackOn Team

(Post-hackathon roadmap)

Phase 5 — Living Flight Graph: Integrate real-time ADS-B, Radar, and Weather feeds to track flights as living entities.
Phase 6 — Predictive Digital Twin: Forecast delays, ETAs, and gate conflicts using XGBoost / LightGBM / Random Forest.
Phase 7 — Autonomous Operations: Auto-create incident tickets, send OCC alerts, and propose schedule/crew/gate adjustments.
Phase 8 — Air Traffic Digital Twin: Simulate airspace, runways, and taxiways to analyze cascade effects between flights.