Why We Built GeoSpy and Why It Became Raven

Before GeoSpy was a product, it was a side project.

My brothers and I were living in a one-bedroom apartment in Boston. I slept on the couch, they split the bed, and after work we spent our nights experimenting with applied AI. What began as a two-week hackathon turned into months of building, testing, breaking things, and trying again.

We did not begin with a business plan for image geolocation. We began with a frustration: some of the most impressive AI research in the world never reaches the people who could actually use it.

That frustration became Graylark. One experiment inside it became GeoSpy. The people who found GeoSpy, and the problems they brought us, eventually turned it into Raven.

Research only matters when it reaches the field

Before Graylark, I worked on applied AI for defense: embedded autonomy, navigation in GPS-denied environments, mission planning, and other difficult problems where software has to interpret the physical world and make useful decisions.

I loved the research. What bothered me was how often great work stopped at the research stage. A paper could be technically impressive and still have no usable interface, no operational workflow, and no path into the hands of a practitioner.

If a breakthrough never leaves the lab, its real-world value is limited.

Graylark started as our attempt to close that gap. We wanted to take difficult AI problems seriously while caring just as much about product design, deployment, and the person who would ultimately use the system.

The experiment that became GeoSpy

The original idea came from a research paper about using machine learning to play GeoGuessr, the game where players infer a location from a street-level image.

The paper was fascinating, but the authors had not published working source code. So we decided to see whether we could build our own version.

The first GeoSpy was simple: upload an image and receive a predicted location with an explanation. It was a research demo, not an investigative platform. I posted it to Reddit expecting a small group of technically curious users to try it.

Instead, GeoSpy reached one million cumulative users in its first three months.

The growth was exciting, but the more important question was why people were using it. We started contacting users and asking them.

Some were playing GeoGuessr. Some were simply curious about what AI could infer from a photograph. But a meaningful group came from law enforcement and government. They were using GeoSpy to generate leads from images connected to real investigations: images with no GPS data, no obvious landmark, and often almost no context.

They showed us that visual geolocation was not just a clever demonstration. It could help investigators narrow a search area, connect online material to the physical world, and act faster when time mattered.

That changed the direction of the company.

From a viral demo to an operational tool

A product used in an investigation has to meet a much higher standard than a public experiment.

The original GeoSpy could estimate a city or region. Investigators needed to move from a broad area to a street, a building, or a specific address. They needed ranked candidates instead of a single unexplained answer. They needed results that an operator could review, corroborate, and document as part of a normal investigative process.

So we went back to the research.

We built proprietary computer-vision models, expanded our data infrastructure, and developed a workflow that could move from regional geo-estimation to street-level targeting. We designed the system to work from the pixels in an image rather than depending on EXIF, GPS coordinates, or other metadata that is frequently missing, stripped, or unreliable.

We also made a deliberate product choice: Raven produces investigative leads, not automatic verdicts. The platform returns ranked candidates and gives the operator the context needed to evaluate them. The final judgment stays with the investigator and should be corroborated through normal procedure.

That distinction matters. Powerful AI should help a trained professional reason more effectively; it should not ask them to surrender judgment to a black box.

Our roadmap comes from the people doing the work

One of the most important lessons from GeoSpy was that the best product ideas often arrive as an offhand question from a user.

During a conversation with investigators at Las Vegas Metropolitan Police, someone asked whether we could identify a vehicle from a difficult image. Not a perfect catalog photograph, but perhaps an interior, a partial body panel, a headrest, a dashboard, or a blurry exterior view.

We took the question back to the research team and built a demonstration. That experiment became Raven's vehicle identification capability, which returns ranked candidates for make, model, generation, and year.

The same pattern has shaped the rest of the platform. Investigators do not experience a case as a collection of separate AI models. They may need to determine where an image was captured, identify the vehicle in it, translate visible text, assess whether the content was manipulated, and organize those findings into one reviewable workflow.

We build around that reality.

Internally, we call some of these experiments "side quests." The phrase is lighthearted, but the principle is serious: a research team needs room to follow difficult questions before the commercial value is obvious. GeoSpy itself began as a side quest. Vehicle identification did too.

The discipline is knowing when an experiment solves a real practitioner problem, and then doing the harder work required to turn it into a dependable product.

Why GeoSpy became Raven

GeoSpy described where we started: AI-powered image geolocation.

It no longer described the full platform.

Raven can help an investigator answer several questions from a single image or video frame:

• Where was this captured? Raven estimates a region and can narrow the result to ranked street-level candidates.
• What vehicle is shown? Raven analyzes partial interior and exterior signals to identify likely make, model, generation, and year.
• Can this image be trusted? Raven can flag signs of synthetic generation or manipulation for further review.

These capabilities belong to a broader category we call frontline visual intelligence: using purpose-built vision models to extract actionable information from imagery that appears to contain very little.

The rename was not a break from GeoSpy. It was the clearest way to describe what GeoSpy had grown into. Raven is built by the same Graylark team, on the same visual-geolocation foundation, with a broader operational mission.

An intelligence layer for the visual world

The internet is becoming more visual. Investigations increasingly begin with a screenshot, a livestream, a marketplace listing, a phone extraction, or a short video clip. The relevant clue may be a road surface, a roofline, the shape of a headrest, the vegetation outside a window, or an inconsistency that suggests an image was generated.

Traditional search was built primarily around text. The next generation of investigative tools must be able to interpret the physical world represented in images and video.

That is the long-term vision for Graylark: build an intelligence layer that can look at visual media and help a practitioner understand what is present, where it came from, whether it is authentic, and how it connects to the rest of a case.

Law enforcement is where we are beginning because the need is immediate and the consequences matter. The people doing this work are routinely asked to make sense of low-context media under time pressure. They deserve tools built around their actual workflow, not generic AI repackaged for a police department.

We built it to be used

GeoSpy began with curiosity. Graylark became a company because practitioners showed us that the technology could matter.

The mission since then has been consistent: take difficult AI research out of the lab and put it into the hands of people doing consequential work, thoughtfully, responsibly, and in a form they can actually use.

The first GeoSpy demo answered one question: Where was this photo taken?

Raven is being built for the questions that come next.

GeoSpy is now Raven, Graylark's frontline visual intelligence platform. Explore Raven's capabilities or read about the evolution from GeoSpy to Raven. Graylark has raised a $10.7 million seed round to continue building it.