The Tech Behind Raven

Ever uploaded a photo to Raven and wondered what happens in the seconds before a location guess pops up on your screen? It can feel a bit like magic. You show it a picture of a street corner, a mountain vista, or a quiet cafe, and it returns a plausible guess, sometimes with startling accuracy.

But it’s not magic, of course. It’s a fascinating, carefully designed system built on some of the most advanced AI technology available today. We wanted to pull back the curtain a bit and give you a candid look under the hood—no marketing fluff, just a straightforward explanation of the tech that makes Raven tick.

The AI That's Seen the World

At the heart of Raven is a powerful tool called a "multimodal AI model." That’s a fancy term for an AI that can understand and process more than one type of information at once. While some AIs just work with text and others just with images, a multimodal model can understand the relationship between them.

Specifically, Raven uses Google's Gemini vision model. Think of Gemini as an incredibly well-traveled, observant friend who has spent years studying a library of billions of images and texts from across the globe. It hasn't "seen" the world in the human sense, but it has processed enough data to learn the subtle patterns that define a place. It can recognize the distinctive terracotta roof tiles of Tuscany, the unique font on a Tokyo street sign, the specific type of acacia tree found in the African savanna, or the tell-tale yellow lane markings of a North American highway.

When you upload a photo, you're not just sending pixels. You're sending a complex tapestry of visual clues. Gemini’s job is to look at that tapestry and say, "Based on everything I've learned, the combination of this style of architecture, that type of vegetation, and the quality of light is most statistically probable in this part of the world."

The Short, Secret Life of Your Photo

This is the part we’re most proud of, and it’s a core principle of how Raven was designed: we never, ever store your photos.

When you upload an image, it takes a very short, direct, and temporary journey. Your browser converts the image into a long string of text (a format called base64), which is sent securely to our server. Our server doesn't save this to a hard drive, a database, or any kind of permanent storage. Instead, it holds it in its active memory—think of it as short-term RAM—just long enough to pass it directly to Google's Gemini API.

Once Gemini analyzes the image and sends its guess back to us, our server immediately discards the image data from its memory. The request is complete, and the photo is gone. It was never written to a disk. It was never saved in a bucket. It's the digital equivalent of showing a friend a photo on your phone, letting them look, and then immediately putting your phone back in your pocket.

This "process-in-memory-and-discard" approach is a deliberate choice. Your photos are your own, and our goal is to provide a fun service without creating a database of user images. We don’t want that responsibility, and we believe you should have that privacy.

A Confident Guess Is Still a Guess

You’ve probably noticed that Raven provides a "confidence" score with its results. It's easy to see "95% confidence" and assume that means it's 95% likely to be correct. But that’s not quite what it means.

The confidence score is the model's own internal measure of certainty. It’s the AI saying, "Based on the patterns in my training data, I am 95% certain that my answer aligns with what I've been taught." It’s a measure of statistical alignment, not a guarantee of real-world accuracy.

Imagine an expert antique dealer looking at a Roman coin. They might say, "I'm 98% confident this is from the reign of Emperor Augustus." Their confidence is based on immense experience, but new information could emerge, or a subtle forgery could fool them. The AI is similar. It can be highly confident and still be completely wrong, especially if a photo contains misleading or ambiguous clues (like a Parisian-style cafe in Las Vegas).

This is why we are crystal clear that Raven is an entertainment and informational tool. It’s a brilliant, highly-educated guesser, but it is not an infallible oracle. It’s for satisfying curiosity, playing a game with your friends on our mobile app Geospy AI, or getting a general idea of where a long-lost family photo might have been taken. It is not, and never will be, a surveillance or intelligence tool. It lacks the precision, the verification, and the access to private data that such tools would require. It's just looking at the pixels you provide.

We built Raven to be a fun, privacy-first peek into the incredible power of modern AI. It’s a testament to how machines can learn to perceive the world in a way that feels almost human. We hope this peek behind the curtain makes your next guess a little more interesting.