Ares | Devpost

Inspiration

Air traffic controllers handle one of the most safety-critical jobs in the world, managing hundreds of aircraft daily with zero margin for error. After learning about the Princeton team's vision for AI-assisted ATC and studying real NTSB incident reports, we realized AI could revolutionize air safety - not by replacing controllers, but by providing them with superhuman awareness and predictive insights. We wanted to build a system that could detect conflicts before they happen, learn from every aviation incident in history, and give controllers the tools to prevent accidents.

What it does

ARES is an autonomous Air Traffic Control system powered by 5 specialized AI agents working in parallel. The Risk Detection Agent scans for separation violations and runway incursions, the Conflict Resolution Agent plans avoidance maneuvers, the Weather Advisory Agent issues weather-related clearances, the Approach Sequencing Agent optimizes arrival flows, and the Voice Communications Agent generates realistic ATC clearances with urgency modulation. The system features a task-based workflow that auto-generates actionable items from alerts with intelligent deduplication, a 6-hour weather forecast timeline showing predicted risk levels, and a conflict predictor that forecasts separation violations at NOW/+5min/+10min/+15min intervals. Controllers see everything on an interactive map with real-time updates, can resolve tasks with one click, and hear AI-generated voice clearances that sound like real controllers.

How we built it

We built a FastAPI backend with WebSocket real-time communication, integrating OpenAI GPT-4 for multi-agent intelligence and ElevenLabs for voice synthesis with dynamic urgency levels. The system uses Redis for caching and pub/sub, connects to the OpenSky Network for live flight tracking, and leverages an NTSB incident database of 80,000+ aviation accidents for risk scoring. The React/TypeScript frontend uses Mapbox GL for map visualization, Tailwind CSS for the ATC-style interface, and custom timeline components for weather and conflict prediction. We implemented a sophisticated task management system with fingerprinting (icao24 + category + priority) for deduplication and auto-expiration after 10 minutes. The multi-agent architecture allows agents to work in parallel, analyzing different risk factors simultaneously and combining their insights for comprehensive situational awareness.

Challenges we ran into

The biggest challenge was getting the multi-agent system to work cohesively without conflicts - we initially had agents making contradictory recommendations. We solved this by implementing a priority hierarchy and conflict resolution logic. Real-time WebSocket updates were crashing under high load, which we fixed by implementing Redis pub/sub and proper async handling. The task deduplication system was tricky - we had to design a fingerprinting algorithm that could recognize duplicate alerts for the same aircraft/risk combination without creating false duplicates. Voice synthesis with dynamic urgency was particularly challenging - we had to map risk scores to voice parameters (speed, pitch, stability) to make critical alerts sound appropriately urgent. Getting CORS to work properly between frontend and backend took longer than expected, and we had to handle embedded git repositories cleanly for deployment.

Accomplishments that we're proud of

We're incredibly proud of building a fully functional autonomous ATC system in a hackathon timeframe with 5 AI agents working in parallel, real-time conflict detection, and predictive analytics. The task system's intelligent deduplication means controllers never see the same alert twice, and auto-expiration ensures their workspace stays clean. The timeline predictors give controllers unprecedented foresight - they can see conflicts 15 minutes before they happen and weather changes 6 hours in advance. The voice synthesis sounds remarkably realistic with appropriate urgency modulation. We successfully integrated multiple complex APIs (OpenAI, ElevenLabs, OpenSky, NTSB database) into a cohesive system. Most importantly, we built something that could genuinely save lives - every feature is designed around real aviation safety needs from actual NTSB incident reports.

What we learned

We learned that multi-agent AI systems require careful orchestration and conflict resolution - having multiple intelligent agents is powerful but complex. Real-time systems need robust error handling and graceful degradation - aviation safety can't tolerate crashes. Task deduplication is harder than it seems and requires domain-specific fingerprinting logic. Voice synthesis can convey urgency through subtle parameter changes (speed, pitch, stability). Redis pub/sub is excellent for real-time updates at scale. The NTSB database is an incredible learning resource - we found patterns in incidents that informed our risk scoring algorithms. We also learned that building for safety-critical systems requires a different mindset - every edge case matters, every timeout needs handling, and comprehensive logging is essential for debugging issues that could affect real operations.

What's next for Ares

We want to integrate real historical incident playback so controllers can train on actual NTSB scenarios and see how ARES would have detected them. Adding machine learning to predict controller workload based on traffic patterns and automatically suggest workload-balancing strategies would be huge. We're planning to implement collaborative multi-sector coordination where ARES agents communicate across sectors to optimize handoffs. Real-time integration with FAA SWIM (System Wide Information Management) for live US airspace data would make this production-ready. Adding natural language voice control so controllers can query the system by speaking would be a game-changer. We also want to build a replay system that records every decision and creates a detailed audit trail for post-incident analysis. Long-term, we envision ARES becoming a training platform for new controllers and a safety analysis tool for aviation authorities to identify systemic risks before they cause accidents.