General aviation / Collision avoidance
Traffic displays only show aircraft that broadcast their position. Atlas finds nearby aircraft visually, from an iPhone camera in the cockpit, so it works whether or not the other aircraft is equipped. No cooperation required from anyone else in the sky.
01 The problem
“See-and-avoid” is still the backbone of GA safety, but human scanning has limits, and the electronic tools meant to help only go so far.
Systems like ADS-B show traffic that’s equipped and broadcasting, which means the aircraft that aren’t stay invisible. Safety today depends on regulations, and on other pilots opting in.
The aircraft you can’t see is the one that matters.
Cooperative only
ADS-B / transponder traffic surfaces. Non-equipped aircraft do not.
Opt-in dependent
Visibility relies on the other aircraft choosing to broadcast.
Human fallback
See-and-avoid carries the gap, limited by attention and blind spots.
02 The solution
Atlas detects aircraft optically, the way a pilot’s eyes would, but it never stops scanning its field of view. An iPhone mounted in the cockpit runs the whole system on the device, targeting detection at 1 to 2 miles. Because it sees rather than listens for a broadcast, it works whether or not the other aircraft is equipped.
This is non-cooperative detection: nothing is required from anyone else in the sky. It’s a supplemental layer that augments the pilot’s scan and existing certified avionics, not a replacement for them.
I
Optical
Detection comes from what the camera sees, not from a cooperating transmitter.
II
Non-cooperative
No transponder, mandate, or opt-in needed from the other aircraft.
III
On-device
Detection runs on an iPhone in the cockpit, with no new certified avionics to install.
03 Flight tests
Our co-founder flew two instrumented test flights as pilot in command, with an iPhone recording 4K video beside an ADS-B receiver we built. Aircraft that broadcast their position label our own footage automatically, with no hand labeling.
On Flight 2, the detector picked out a real aircraft against forested terrain, on video it had never seen.
Flight 2 · Oct 2026
Trained only on public data, the detector locked onto an aircraft as the crew spotted it and held the track for the full encounter, filtering about 300 candidates per frame. Across the clip, the aircraft ranked first of 3,646 candidate tracks.
Flight 1 · Oct 5, 2026
The capture app first compiled on September 27 and flew on October 5. The flight validated the camera, receiver, synchronization and labeling pipeline together.
Ground test · Eugene
Between flights we test against aircraft overhead. Frame by frame, a rooftop or a patch of haze can look as much like an aircraft as the aircraft does. Following every candidate over time is what singles it out: here the aircraft ranked first of 1,296 candidate tracks.
Where this stands: detection is demonstrated on real flight footage, from a small number of encounters. Cutting false alerts and validating across many more flights are the active work.
04 Why this works / why now
The insight
Every cooperative system shares one assumption: the other aircraft is broadcasting. Remove that assumption and the only detection method left standing is seeing the aircraft directly. That’s the wedge, and it doesn’t depend on anyone else acting.
The timing
A phone pilots already carry now has a 4K camera and a GPU strong enough to run our full detection pipeline. We measured it on an iPhone 17 Pro. Optical detection that once needed dedicated hardware is now within reach of a small, focused team.
05 Technology
A distant aircraft is a handful of pixels against bright, changing sky and terrain, seen from a platform that is itself moving and vibrating. It has to be found early enough to matter, on hardware that fits in a pilot’s hand.
Atlas is trained on Amazon’s public Airborne Object Tracking dataset, with our own ADS-B-labeled flight footage added as it accumulates. Every flight we fly makes the training data more like the cockpit it serves.
Find
Scans every 4K frame for small objects that stand out from sky or terrain, bright or dark, and narrows millions of pixels to a few hundred candidates.
Classify
A compact neural network, 1.7 million parameters, scores each candidate as aircraft, bird or clutter.
Track
Follows every candidate across frames, correcting for our own aircraft’s motion, and alerts only when the evidence builds over time.
Label
ADS-B positions are projected into the camera image to label flight footage automatically, with a 5.9-pixel median label error in testing.
Measured on iPhone 17 Pro · Oct 2026
Every stage is ported to the iPhone, and the detection front end is verified to match the desktop reference exactly. It runs at about 10 frames per second on the live camera, and we have measured where the remaining time goes.
10.1 fps today
06 Progress
Each step was built, flown and measured before the next one started.
The full motivation and plan: Thesis prospectus →
2024
Trained in Cessna 172s and the DA-20. Atlas started with a near-miss.
Jan 2025
Started as a Clark Honors College thesis at the University of Oregon.
2026
$5,000 award. Requirements gathered from 12+ pilots and instructors and 2 partner flight schools.
Sep 2026
An iPhone app that records 4K video, attitude, GPS and pressure altitude in sync with a team-built Stratux ADS-B receiver (about $240 in parts).
Oct 2026
82 minutes and 148,409 frames recorded with the ADS-B link up the whole flight. ADS-B positions now label footage automatically.
Oct 2026
Trained only on public data, the detector locked onto an aircraft as the crew spotted it and held the track for the full 17-second encounter.
Oct 2026
The full detection pipeline ported to an iPhone 17 Pro and verified exact against the desktop reference. 10 fps on the live 4K camera.
Fall 2026
Selected for the accelerator’s fall 2026 cohort.
Next
Pre-briefed passes with a known aircraft so every flight yields labeled training data from our own camera, plus a faster pipeline toward 24 fps and fewer false alerts.
07 Recognition
Accelerator
Oregon UAS Accelerator
Cohort 4 · Fall 2026
Award
Oregon Innovation Challenge
Fellows · $5,000
Research
Clark Honors College thesis
University of Oregon
Customer discovery
12+ pilots and instructors
2 partner flight schools
08 Team
Founder-market fit is the point: a private pilot who leads the detection model and flies its tests. The problem is understood from inside the cockpit, and the product is checked there too.
Co-founder · CEO · ML Lead · Test Pilot
Private pilot with 80 hours in Cessna 172s and the DA-20. Leads the machine-learning work on the detection system and flew both instrumented flight tests as pilot in command. Computer science at the University of Oregon’s Clark Honors College.
Co-founder · CTO
Leads development of the Atlas app. Software engineer with years of experience in startups, including running his own software consulting business.
09 Mission
Put more safety in the hands of the pilots flying the airplane, without waiting on anyone else to act.
10 Contact
Investors, technical builders, pilots and flight schools. If the wedge resonates, we’d like to talk.
Prefer email? kellen@fly-atlas.com
Message received
Thanks for reaching out. This goes straight to the founders, and we’ll reply directly.
