The pipeline
Anyone can name the food. The hard part is how much.
kcalc.ai separates a plate into its ingredients, estimates each one on its own, and uses the depth sensor in your iPhone so the portion rests on a scale reference rather than a guess.
CaptureLiDAR optional
Three ways in
Photograph
A photo is enough, on any supported iPhone.
LiDAR orbit
On an iPhone with a LiDAR sensor you can orbit the plate instead, and the app captures depth as you move.
Barcode, text and recipe
A barcode resolves against the product database. A typed sentence, a pasted recipe or a recipe link all work too.
Scale referenceTwo metric bars
The table sets the scale
During a scan the app isolates the food from the table, fits the real table surface, and burns two metric scale bars onto a copy of the photo. The model reads the portion against a true scale reference instead of an assumption about how big a bowl is.
SeparationOne portion per ingredient
Separate, then estimate
A mixed plate is not one food. The pipeline pulls it apart into the things that are actually there, estimates the portion of each, and adds them back up. A known packaged product skips the guessing entirely and takes its figures from the label.
The memory3.2 million foods
A large memory of food
Behind the estimate is a knowledge base of about 3.2 million products, dishes and ingredients, with nutrition data and portion knowledge to draw on.
Your context
A short memory of you
What goes to the model is what it needs for this one estimate: the photo, the depth scan, your note and any answer you give. When the estimate is done, that context does not carry over to the next one.
Your meal and its result are saved to your account so you can review and correct them; that is your record, not the model's memory.
The waitSeconds to minutes
It takes a moment, on purpose
The job is queued on our servers and takes anywhere from a few seconds to several minutes, depending on what is in front of it. You can watch it work, or close the app and let the notification find you.
The inboxOptional
One question, sometimes
Where the answer turns on something the photo cannot settle, the app asks rather than guessing. You can leave it: the app waits, then finishes with its own best estimate.
ResidencyStored in Sydney
Where it runs
Most of the AI runs on our own GPU servers in Australia. When they are unavailable, the meal's photos and text go to Anthropic in the United States. Your photos and scans are stored in Sydney and age out of storage on a schedule. The retention periods, and the full list of providers including the search index hosted overseas, are in the privacy policy.
Read the privacy policyTraining
It does not remember you
Your data is not training material. It is context, handed to the model for one estimate and nothing else.
Estimates
Estimates, honestly
These are estimates, not measurements. The model can be wrong, sometimes by a lot, especially for food it cannot see clearly. Every figure is editable, and your corrections stay with your meal.
Questions
What people ask first
Does kcalc.ai need the internet?
Yes, for the estimate itself: the AI runs on our servers, not on the phone. The app is offline-first once you are signed in, and a meal captured without signal is saved, then estimated when the connection returns.
Where does my meal data go?
Your account, meals, photos and depth scans are stored in Sydney. Two things go to the United States: a search index receives search data such as an embedding of the food description, and when our own AI servers are unavailable, Anthropic receives the meal's photos and the text sent with them to make the estimate. The privacy policy names every provider.
Is the AI trained on my photos?
No. Your data is not training material. It is context, handed to the model for one estimate and nothing else.
The companyHex Pro
Why we built it this way
Hex Pro is a small Australian software company. We store data in Australia, we run no third-party advertising and we do not harvest.
Read the Hex Pro missionWhere nextTwo pages / one pipeline