AirPilot: Your Wings, Our Watch Why I Built It Airplane crashes haunt us all—a shadow that lingers during every flight. The Ahmedabad tragedy on June 12, 2025, where a Boeing 787 plummeted into a hostel, killing 279, hit me hard. It wasn’t just news; it was a wake-up call. With 80% of crashes tied to human error, I asked: What if we could catch those mistakes before they spiral?

What’s AirPilot? AirPilot is your AI co-pilot, always alert. It scans live flight data—altitude, speed, warnings—and matches it to past crashes using MongoDB’s vector search. Powered by LangChain, it whispers clear, lifesaving advice: “Climb 500 ft,” “Throttle up to 70%,” or “Abort takeoff.” It’s a guardian for pilots, keeping skies safe.

Tools I Used Backend: FastAPI, LangChain, Google Vertex AI Database: MongoDB Atlas on Google Cloud Frontend: React, ElevenLabs, WebSpeech Hosting: Render AI: Lang chain with Google Gemma, Hugging Face Sentence Transformer

How I Built It Demo Flights: I crafted six crash scenarios—Korean Air 801, Asiana 214, Air France 447, Colgan 3407, Turkish 1951, Tenerife KLM 4805—as JSONs, pinning their fatal moments (e.g., Colgan’s 131-knot stall). MongoDB Setup: Loaded JSONs into MongoDB Atlas using pymongo, ready for vector search. Vector Embeddings: Used Vertex AI to turn telemetry (altitude, speed, warnings) into 768D vectors, stored in MongoDB. LangChain Magic: Built three chains—emergency advice, risk explanation, turbulence chat—tuned for precise, TCAS-compliant responses. Testing: Simulated flights via FastAPI, testing advice against real-world standards, Included test.py files for each simulated flights to test different scenarios Frontend Flair: Inspired by X-Plane dashboards, I used React or a clean UI, and also, OpenSKY api for live flight data

Challenges Query Routing: Pilot queries were messy. I built a LangChain router to direct them to the right chain. Hallucinations: AI invented crash details. Fixed it by grounding prompts with specific JSON data. Vector Search: MongoDB embeddings misfired until I standardized dimensions and tested small-scale.

What I Learned Data Drives Trust: Clean, structured JSONs kill AI drift. Modularity Wins: Separate chains and endpoints make scaling easy. Pilots Need Clarity: Simple, urgent advice builds confidence.

Improvements More Simulations: Add 50+ crashes and edge cases via Monte Carlo methods. Live Weather: Stream METAR data for better turbulence advice. System Health: Track hydraulic pressure, fuel flow to catch mechanical failures. Collision Detection: Add nearby_aircraft data for TCAS-like alerts.

What’s Next? Public demo with pilot feedback. Integration with flight sims like Microsoft Flight Simulator. AR cockpit alerts for real-time guidance. Certification for commercial cockpits.

Tech Stack Frontend: React JS, Elevenlabs and Webspeech for TTS Backend: FastAPI, LangChain Database: MongoDB Atlas AI: Google Vertex AI, Google Gemma via Ollama

What if the next crash never happens? That’s AirPilot’s promise. — Miraj Prasain

Built With

Share this project:

Updates

posted an update

The backend chat disconnection issue that occurred during the original submission has been fully resolved. The bug was related to deployment instability on Render, which caused intermittent failures in the real-time chat response system.

This fix is not reflected in the current Devpost demo due to submission time constraints and deployment lag. However, the updated version is working as intended and can be shared upon request.

Please feel free to reach out if you'd like access to the fixed version or a live walkthrough.

— Miraz Prasain

Log in or sign up for Devpost to join the conversation.