AI Engineer NYC 2026
Join AI Engineer and Major League Hacking (MLH) on Sunday, October 11, for a full day of building, experimenting, and meeting fellow AI engineers.
Landed
AI teammate for NYC’s helicopter desk: handles routine radio calls with Gemini and warns controllers a minute before a runway conflict.
About the project
Inspiration
On January 29, 2025, an Army helicopter and PSA Airlines flight 5342 collided near Reagan National. 67 people died. The NTSB found one controller working the combined local and helicopter positions, and cited “the lack of safety alerts.” The conflict alert sounded only 26 seconds before impact. In New York, helicopters on the Hudson and East River routes self-announce on frequencies no controller is required to monitor, right next to LaGuardia’s approaches.
What it does
Landed listens to helicopter radio calls and tracks every aircraft by ADS-B. On a simulated desk, it handles routine medevac check-ins using a fixed list of approved phrases, and hands anything risky to the human controller with a reason and a countdown. In our demo, a medevac at 600 ft heads toward the LGA runway 22 final, and Landed alerts the controller about 84 seconds before it enters the protected zone.
How we built it
- Gemini for speech to text, parsing radio calls into structured data, function calling, and text-to-speech for the agent’s voice
- A conflict engine that projects each helicopter’s path and checks it against LGA approach and departure zones, built from FAA procedure data
- Real ADS-B replay from adsb.lol (17 helicopters, Sat Oct 10, 11:00 to 11:15 ET)
- React and MapLibre for the controller display, FastAPI for the agent server
Safety by design
The agent only speaks approved phrases, never clears anyone toward a runway, and hands off on emergencies, low confidence, confusable callsigns, or any request outside its list. A false alarm costs the controller a second to dismiss. A miss is today’s status quo. Simulation only: Landed never transmits on a real frequency.
Results
- 0 false approach or departure handoffs across 17 real helicopters in the demo window
- Gemini parses a radio call in 0.89 s median, with 100% valid output on 50 of 50 test calls
- Handing off below 0.8 confidence catches 92.9% of parse errors
- All 6 scripted scenarios pass
Challenges
Real ATC audio is noisy (33% word error rate on real tower audio), so we lean on confidence-based handoffs and ADS-B as an independent check.
What’s next
Listen-only alerts first, the same decision-support role the FAA already uses AI for. Then a trial at a quiet field or a medevac-only route, and fine-tuning on tower audio.
Landed is designed to raise an earlier alert, but nobody has tested what would have happened at DCA.
ADS-B data: adsb.lol, ODbL 1.0. Map data: OpenStreetMap contributors, ODbL.
Built with
AI tools
Prize Categories Entered
- Submitted
Best Use of Gemini API
Gemini runs Landed's whole voice pipeline. It transcribes helicopter radio calls, parses each one into structured data (callsign, intent, altitude) with a confidence score and a reason, picks the agent's reply from an approved phrase list through function calling, and speaks it with Gemini text-to-speech. It parses a call in 0.89 s median, with 100% valid output on our 50-call test set, and handing off below 0.8 confidence catches 92.9% of parse errors.
- Event prizeSubmitted
Best Overall
Landed takes on a real, deadly problem. The NTSB found "the lack of safety alerts" behind the 2025 DCA midair collision, where the alert came only 26 seconds before impact. Landed handles routine medevac helicopter calls and warns the controller about 84 seconds before a helicopter enters an airliner's approach path. It runs end to end on real NYC ADS-B data with 0 false handoffs across 17 real helicopters, and it is safe by design: approved phrases only, a handoff whenever it's unsure, and simulation only.