The Tech Behind Raven

Have you ever uploaded a photograph to Raven and wondered what happens in the seconds before a guess appears? It can feel like a conjuring trick. You show it a street corner, a mountain ridge or a quiet cafe, and back comes a plausible place, sometimes with unsettling accuracy.

It is not a trick, and the machinery is smaller than most people assume. There is no index of the world's photographs, no reverse image search and no hidden metadata harvest. There is one model, one request, and a deliberate decision about what not to keep. Here is the whole of it, without marketing varnish.

The flow is short enough to describe in five steps, and knowing them explains most of the product's behaviour, including its limits.

1. 1.You upload. The file goes from your browser to Raven's server over an encrypted connection. Nothing is analysed in the browser itself.
2. 2.The server checks the file. Size is capped, and the file type is confirmed by reading the opening bytes rather than by trusting the label the browser attached to it.
3. 3.The image is encoded in memory. The bytes are turned into text so they can travel inside an ordinary API request, and that text exists only as a variable in a running function.
4. 4.Gemini reads it. The encoded picture and a written instruction go to Google's model in one call, and it replies with a place, a short explanation and a confidence figure.
5. 5.The answer is parsed and returned, and the image is released. The response is turned into structured data for the interface, the request ends, and the photograph goes with it.

Two details in that list are worth pausing on. The first is that the type check reads the file's opening bytes, which is the only honest way to know what a file is: a JPEG begins with a fixed three-byte signature no matter what extension somebody typed, and the table of file signatures is the reference for this. Trusting the content type the browser announces would mean trusting whoever wrote the request.

The second is the encoding step. Images travel to the model as base64 text, which expands the payload by roughly 33 per cent because every three bytes become four characters. That overhead is the real reason there is an upload size limit, and it is also a useful reminder of how little ceremony is involved: your photograph spends its entire life in this system as a string in memory.

Think of Gemini as an extremely well-travelled, extremely well-read observer that has never actually been anywhere. It has processed enough material to have absorbed the patterns that define places: the terracotta pantiles of Tuscany, the type on a Tokyo street sign, the flat-topped acacia of the East African savanna, the yellow centre line of a North American highway. None of that is stored as a list of facts to look up. It is a set of statistical regularities that fire, or fail to fire, when a new picture arrives.

The reason a single model can weigh a roofline against a plant against an alphabet is that it holds images and language in one shared representation rather than running separate detectors and stapling the outputs together. That property, and its consequences, are the subject of how an image model reads a photograph. It also explains why the reasoning shifts so much with the scene: a dense street and an empty valley hand the model completely different toolkits, as urban and rural scenes shows.

This is the part of the design worth being most explicit about, because it is easy to claim and easy to quietly break. Your browser sends the file, the server holds it in active memory just long enough to hand it to the model, the model answers, and the request completes. The photograph was never written to a disk, never saved in a bucket, never recorded in a table. It is the digital equivalent of showing a friend a picture on your phone, letting them look, and putting the phone back in your pocket.

It is worth adding what is not read as well as what is not kept. Camera files carry an EXIF block, a metadata standard in use since 1995, and on a phone that block often contains precise coordinates. Raven's analysis does not consult it. The guess is made from visible evidence only, which is why a screenshot with every tag stripped produces the same kind of answer as the original file. If the pixels do not show it, the model does not know it.

It is tempting to read 95 per cent as a nine-in-ten guarantee. That is not the claim. The figure describes the shape of the model's own preference, and nothing in the pipeline verifies the result against the world. Consider an antiques dealer saying they are almost certain a coin dates from a particular reign: the confidence rests on long experience, and long experience can still be fooled by a good forgery. A model can be emphatic and wrong in exactly the same way, most often when a scene resembles somewhere it has seen a great deal of, such as a Parisian-style cafe front that happens to be in Las Vegas.

There is a second reason to hold the number loosely. Human guessers and models fail differently: people lean on where they have personally been, while a model leans on how often a place has been photographed, which is a bias with its own geography. That divergence is examined in why humans and AI guess locations differently, and it shows up most clearly on near-identical neighbours, the case explored in telling neighbouring countries apart.

The economics are unglamorous and worth stating plainly. Every guess costs something to run, so the analysis feature sits behind a single sign-in and the web version includes one free guess per account. Everything else here, including this article, is open. When the free guess is used up, the honest recommendation is the phone: Geospy AI is free on iOS and runs the same kind of reasoning against your camera roll.

A short list of absences is often more informative than a feature list. Raven does not perform reverse image search against the web. It does not match faces or attempt to identify anybody. It does not read location metadata. It does not build a history of what you have uploaded, because it keeps no uploads to build one from. And it does not present its answers as findings; the interface shows a confidence indicator precisely because the output is an estimate.

What is left is a narrow and genuinely entertaining thing: a model looking at pixels and reasoning aloud about them. Most of the pleasure is in the reasoning rather than the pin, especially when the clue turns out to be something you would never have noticed, such as the particular scrub at the roadside that pins a climate band, as how vegetation reveals climate explains.

We built Raven as a small, privacy-first look at what modern models can do with an ordinary photograph. It is a demonstration that machines have learned to perceive in a way that feels almost human, and a reminder that almost is doing a lot of work in that sentence.