How we built accurate food recognition by testing against thousands of verified nutrition database entries and learning from real user corrections.
Standard vision models can identify a "Chicken Breast" but often miss the cooking oils, butter, or sauces added during preparation. A pan-seared chicken breast can have significantly more calories than a raw one.
Accounts for preparation methods and cooking additions
No AI is perfect out of the box. We built a system that continuously improves by learning from user corrections and feedback.
Advanced vision models analyze the photo and provide an initial estimate of calories and macros.
If our estimate is off, users can correct it. We learn from every correction to improve accuracy.
Corrections are stored in our Smart Vault, so similar images get better estimates over time.
A handful of almonds (high calorie density) can look similar to a handful of grapes (low calorie density) in a photo. Accurate portion estimation is crucial.
Our Approach
We use reference objects in the image (plates, utensils, hands) to estimate portion sizes more accurately. Our system has been tested against thousands of verified nutrition database entries to improve portion recognition.
Validation
Every estimate is validated against professional nutrition databases. We've tested our system with over 5,000 verified meal photos to ensure accuracy.
We've tested our system against thousands of meals from professional nutrition databases, comparing our AI estimates to verified calorie and macro values.
Our Smart Vault system learns from every user correction, building a knowledge base that improves accuracy over time.