How Doy reads your plate: a look inside the AI food scan

What happens when you scan a plate in Doydum?

The short version fits in one breath: you take a photo, Doy reads it, you check his work. The longer version has six steps, and each one is there for a reason.

1. Shoot. You take the photo in Doydum’s own camera.
2. Crop. You trim the frame down to the plate.
3. Hint. If you like, you add a few words about what the camera can’t see.
4. Estimate. The photo goes to an AI model, Google’s Gemini (or Anthropic’s Claude, if Gemini can’t answer), which names each item with grams, calories and macros, plus a confidence.
5. Review. You see every item, fix what’s off and remove what isn’t there.
6. Save. The meal joins your day, its photo joins your gallery, and a row is planted on Doy’s farm.

You can scan up to 30 plates a day. Typing a food from the list or copying a meal from your gallery doesn’t use a scan. There’s a quick tour of the scan on our homepage, too.

Why does Doydum use its own camera, a crop and a hint?

We built our own camera because the moment you take the photo matters more than it looks. A hurried shot that’s half table, half phone screen and a corner of someone else’s plate hands the model a puzzle before it has even started. In Doydum’s camera you frame the plate and shoot, and the shot goes straight to the crop.

The crop is the quiet hero. Trimming away the tablecloth, the bread basket that isn’t yours and your neighbour’s glass tells the model exactly which food to read. It also keeps the photo saved with your meal tidy. Before it leaves your phone, the photo is shrunk to a compressed JPEG of at most 1,280 pixels: plenty for reading a plate, and much lighter to send.

The hint is optional, a few words at most, and it can matter a lot. A photo shows the surface of a meal; a hint tells the model what the surface hides:

• “Cooked in butter” or “fried in olive oil”
• “Two sugars in the tea”
• “Half portion” or “I ate about two thirds”
• “Homemade, lean mince”
• “The filling is cheese and spinach”

The model reads your hint together with the photo, so there’s no special format to learn. Write it the way you’d tell a friend.

What does the AI send back?

The model doesn’t return one number for the whole plate. It names each item it recognises and gives it grams, calories, protein, carbs, fat and a confidence, plus a name for the plate, the way you’d say it at the table. That item-by-item list is what makes the estimate checkable: you can see exactly where every number comes from.

Here’s an illustrative example of what a simple lunch could look like after a scan. The values come from our calorie chart; a real scan sizes the portions from your photo.

Every line also carries a confidence. When the overall confidence is low, Doy tells you before you save, so you know which lines deserve a second look. Answers come back in the app’s language, English or Turkish.

What does Doy do while the estimate is on its way?

An estimate takes a little while, and we didn’t want a spinner. So Doy does the reading, in pixels, at a little table:

1. He brings a phone up from behind the table’s edge.
2. He switches on the flashlight, squints one eye and takes aim.
3. The shutter flashes and he blinks.
4. A scan line sweeps across the plate while he reads, wide-eyed, one brow raised.
5. When the numbers are ready, the macros fly into the phone.
6. Then he reacts to the dish.

None of these is a judgement. A flex isn’t “better” than a zap; they’re just Doy’s way of saying he saw what you’re eating. With Reduce Motion turned on in your iPhone’s settings, he keeps to still poses instead of the whole show.

How do you check and edit the estimate?

The estimate is a draft until you say otherwise. On the review screen every item is yours to change:

• Change the grams. Calories and macros follow the amount, so fixing a portion fixes everything on that line.
• Remove what isn’t there. If the model spotted a lemon wedge you didn’t eat, take it off.
• Rename the plate. The model suggests a name, such as “Lentil soup with bread”; change it to whatever you’ll recognise later.
• Log what it missed. Anything the photo couldn’t show, like the ayran you drank before you took it, you can log from the food list.

When you save, Doy takes his bite, the meal joins your day, and a row is planted in today’s bed on Doy’s farm. Our guide on how accurate AI calorie counters are goes deeper.

What can a photo estimate see, and what can’t it?

We’d rather you trust the scan for the right reasons, so here’s the honest picture. A model reading a photo is good at some things and simply blind to others.

The same dish also varies from kitchen to kitchen. A lahmacun can be small and thin or large and generous; a homemade lentil soup may or may not have butter swirled on top; in mantı with yogurt, a photo can’t count the dumplings under the yogurt. The model makes a reasonable guess, which is why your review and your hint matter. Treat the result as a careful estimate you can edit, never as an exact count.

For packaged food, the barcode is the better tool: it reads the product’s own label data from Open Food Facts. Our barcode guide explains how that works.

What happens to your plate’s photo?

It’s your meal, so here is exactly where the photo goes:

• It’s sent to Google’s Gemini (or to Anthropic’s Claude, if Gemini can’t answer) only for the estimate you asked for.
• If you save the meal, the photo is kept with it, so your meal gallery can show every plate, newest first.
• If you delete the meal, its photo goes with it (unless a copy of that meal on another day still shows it).
• If a scan fails, or you leave the review without saving, the photo is deleted.
• When you delete your account from Settings, your photos are erased first.

There are no ads and no tracking SDK in Doydum, and you can export all your data as JSON from Settings. The full story is on our privacy page.

How do you take a photo the AI can read well?

A few small habits make the model’s job easier:

• Shoot before the first bite. A full plate is easier to read than a half-eaten one.
• Find some light. Daylight is best; in a dim room, move the plate towards the light.
• Get the whole plate in, then crop. The edges of the plate help with judging size.
• Tilt a little. Straight down works for flat plates; a slight angle shows the depth of a bowl or a glass.
• Keep the cutlery in. A fork next to the plate gives a sense of scale.
• One meal per scan. Two people’s plates in one photo make it hard to tell whose is whose.
• Say what’s hidden. Oil, butter, sugar, fillings and portion size: a short hint covers them all.

And if a plate just won’t scan well, that’s fine. Search the food list or copy a similar meal from your gallery, then adjust the grams. Our guide to the AI calorie counter has more on getting the most out of photo logging.

Questions people ask