Ford Freyberg started his career looking at forests. Specifically, the forests of Idaho, where he spent an internship with NASA studying wildfire susceptibility and risk. It was the kind of work that teaches you to read a landscape from above, to understand how environmental factors interact across thousands of square miles. He didn’t know it at the time, but that internship would set the trajectory for everything that followed.
Today, Freyberg is a product manager at Vexcel, where he helps oversee software called Custom Elements—part of the broader Vexcel Intelligence. The premise is simple: make the real world searchable. Vexcel maintains an aerial imagery library spanning more than 45 countries and territories across six continents, captured at a resolution of roughly 7.5 centimeters, or about the size of a Post-it Note. Custom Elements lets users search that library the way you might search a database—not with keywords on a page, but with natural-language descriptions of physical objects.
No Model Training Required
The experience from the user’s side is strikingly straightforward. There’s no model to train, no dataset to label, and no machine learning pipeline to configure. Users simply type what they are looking for, and the system goes to work, scanning across the imagery library for matches. Behind the scenes, the object detection engine processes the request against Vexcel’s collection of both ortho (top-down) and oblique (45-degree angle) imagery. That oblique perspective is something most satellite and aerial providers don’t offer, and it matters more than you might think. A propane tank viewed from straight above is a circle. Viewed from an angle, it reveals its shape, its shadow, its relationship to the surrounding terrain. The ability to run detection on that richer perspective opens up possibilities that flat imagery alone can’t match.
Of course, an open-ended system like this comes with an inherent tension: if you can search for anything, you can’t have everything pre-benchmarked. Freyberg is candid about this. Accuracy depends on the object. Large items are readily detectable, but there’s a physical ceiling that’s set by the imagery itself. At 7.5 centimeters of resolution, you can find a car, but you can’t find an ant. That’s not a software limitation; it’s the physics of what a camera can resolve from the altitude at which Vexcel flies.
A New Kind of Workflow
For the people who use this technology day to day, the implications go well beyond convenience. Imagine an analyst at an insurance company tasked with assessing risk across a portfolio of 50,000 properties. Under the old model of manual inspection, that analyst would sit at a desk and pull up image after image, scanning each one by eye, looking for the things that matter: a swimming pool without a fence, a propane tank too close to a structure, or a roof in disrepair. It’s tedious, slow, and fundamentally unscalable. At 50,000 locations, it becomes impossible.
Custom Elements changes the shape of that work. Instead of spending their hours visually sweeping through imagery, the analyst lets the AI handle the search. The system surfaces the relevant locations, and the analyst shifts into a different role, one that plays to human strengths. They triage, prioritize, and make judgment calls. The human is still very much in the loop, but the loop has gotten smaller and more focused. The work moves from finding things to deciding what to do about them.
One question that naturally arises is how current is the data? Vexcel runs consistent capture programs across all its coverage areas. In U.S. urban locations, imagery is refreshed at least twice a year. However, since the detection system analyzes imagery as it was captured on a specific date, the results are a snapshot in time. If a propane tank was removed three months after a flyover, the system would still show it in that image, because that’s what the camera saw. It’s an honest representation of a moment, not a live feed.
What’s Next
Looking ahead, Freyberg is careful to draw a line between what exists today and what’s coming. On-demand object detection across ortho and oblique imagery, with no model training required and global coverage, is available now. Additional capabilities are in development, designed to integrate directly with Custom Elements, but Freyberg is clear about the distinction that what’s shipping and what’s on the roadmap are two different conversations.
If there’s one thing he wants people to understand, it’s that the quality of the AI is only as good as the quality of the imagery underneath it. AI scientists have a phrase for it: “garbage in, garbage out.” Vexcel’s advantage is the inverse. The resolution, the positional accuracy, and the consistency of the capture specs—all of it adds up to the most reliable possible foundation for the detection capabilities built on top. The imagery is the differentiator, and everything else flows from that.
It’s a long way from the forests of Idaho, but in some sense, Freyberg is still doing the same work of reading landscapes from above, trying to understand what’s out there. The difference is that now, with Custom Elements, he’s building the tools that let anyone do it at scale.
