NaventraAI air traffic control console that grades itself against reality.
01
Where the idea came from
I have always been pulled toward aviation, and air traffic control is the part of it that reads like a live systems problem: dozens of aircraft, hard separation rules, weather changing under you, and no room to be wrong. I wanted to see whether I could build a console that did the controller's job on real traffic, not a toy simulation, and then hold it accountable the way nothing usually holds an AI accountable, by checking its calls against what actually happened. Naventra is that experiment turned into a product.
02
Working real traffic
Naventra ingests live ADS-B transponder returns around any of 12 major world airports and the current METAR, then runs an autonomous decision core that does what a TRACON controller does, continuously. It computes head and crosswind components for every runway end and picks the active configuration, classifies each track into a flight phase and builds the arrival sequence by ETA across parallel runways, predicts closest-point-of-approach against terminal minima with a 150 second lookahead, assigns real gates from the airport's actual stand layout, and voices every decision as realistic VHF phraseology with airline callsigns and pilot readbacks. If every live feed is unreachable, a physics-based simulation seeded with the real runway geometry and weather takes over and everything runs identically.
03
Grading itself against reality
The headline feature is the scorecard. When an inbound flight commits to the approach, the core's plan is locked as a prediction: runway end, touchdown ETA, next-to-land order. When the flight actually lands, ground truth is derived purely from observed data, and every locked item is graded against it. Only live traffic banks into the persistent score, go-arounds void their predictions, and unclassifiable landings are discarded rather than guessed. On top sits an online learning layer that updates per-airport runway priors and an ETA bias correction from every verified landing, so the system measurably gets more accurate the longer it watches an airport.
04
The scope
The default view is a 3D TRACON scope built in three.js: drag to orbit, scroll to zoom, with every aircraft at its true altitude, a stem to the ground plane, history trails, heading cones, and extended final-approach centerlines. A 2D toggle gives the classic top-down phosphor scope. Positions are dead-reckoned between polls so motion stays continuous at 60fps, and a built-in Operator's Guide explains every panel and all the ATC terminology for visitors who have never watched a scope before.