How Renzy works: nutrition analysis from a photo

Renzy combines computer vision, a 500+ food database (Spanish + international + packaged products), and multimodal language models to turn a photo into a detailed nutrition breakdown. This is the step-by-step guide — from opening the camera to seeing your calories, protein, carbs, fat, vitamins and a health score.

The 5-step Renzy scan

1. 1Snap a photo of your plate

   Open Renzy, tap the green button and frame the plate. Perfect angle is not required: the AI tolerates odd angles, low light and mixed dishes. Before upload, the photo is scaled down to at most 1,568 pixels on its long side so it travels faster.

2. 2AI identifies the foods

   A multimodal model (vision + language) detects each food in the photo, estimates the portion, and returns a confidence score. It usually gets common dishes right (paella, Caesar salad, omelette, pizza); in complex preparations it picks out the main ingredients and, if it is unsure about one, asks you what it is.

3. 3We cross-reference the nutrition database

   For each food, the model estimates macros (protein, carbs, fat) and micros (sodium, fiber, vitamins A, C and D, iron, calcium). Those figures are checked against our reference database (from USDA and from BEDCA), and the ones that are far off get corrected. For packaged goods, the barcode brings in the values on their label (Open Food Facts).

4. 4We compute the health score

   Every plate gets a score from 1 to 10, worked out the same way for everyone (your goal does not change it): protein, fiber, fruit, vegetables and legumes add points; sodium, free sugar, saturated fat and processed meat take them away, using WHO and AHA thresholds. Below it you see which change would raise it.

5. 5Save or adjust

   If something is off — larger portion, missing ingredient — fix it right there. Save it to your log and see what is left of your day. If you connect Apple Health (iOS) or Health Connect (Android), Renzy reads your weight from it; meals are not written there.

What's behind the AI

No magic. We use a general-purpose AI model (Claude, by Anthropic) that we did not train ourselves: we send it the photo with our instructions, and it tells us which foods it sees and how many grams it estimates. Those figures are checked against our reference food database (USDA and BEDCA values), and the ones that are far off get corrected. A second model, from Google, looks at the same photo on its own and, if it answers in time and the two disagree, confidence drops and the portions get a second pass. It also uses your corrections: if you change a name or a portion, your next photos take it into account. Portions are estimated from visual references (utensils, plate size) and context. When confidence is low, we ask you to confirm: better to ask than to guess wrong.

How accurate is it?

It depends on the dish. With a barcode, the values are the ones on the product label. In a photo, the portion is the hardest part, and for homemade dishes with ingredients you can’t see (oil, sauces) the margin grows. We don’t quote an accuracy figure from memory: we publish the error we measure, kept up to date, on our accuracy page. That’s why every analysis comes with a confidence level and you can correct it before saving.

Privacy: what we do with your photos

Photos are stored encrypted in the cloud and tied to your account, so you can browse your history. To analyse it, the photo is sent to the AI models (Claude and the second model, by Google); we never sell your photos or hand them to anyone else. They only enter our training set if you switch that on in Settings; it is off by default. You can delete your account from Settings (type DELETE to confirm), and your data and photos go with it.

Frequently asked questions