How AI plate analysis works
What the model does, where it's accurate, where it isn't, and how to get the best result.
title: "How AI plate analysis works" description: "What the model does, where it's accurate, where it isn't, and how to get the best result." category: "tracking" order: 2
What happens when you snap
- The photo is sent to Claude (vision-capable model) along with your dietary tags as context.
- The model identifies the dish + components, estimates portions in grams, and pulls per-component macros from our nutrition DB (USDA + Edamam).
- We compute a per-photo confidence score (0–100). Under 60% we ask you to confirm; over 80% we save without nagging.
- You can edit any field before saving.
Where it's accurate
- High accuracy (within 10–15% of dietitian estimate): clearly framed plates with familiar dishes, overhead angle, good light.
- Decent (within 20%): bowls with mixed ingredients, sandwiches, plates with 2–3 components.
Where it struggles
- Crowded or stacked plates — the model can't see what's underneath.
- Liquid volumes — soups, smoothies. We ask for portion confirmation.
- Sauces and dressings — often invisible in photos. Add them manually.
- Multiple plates in frame — capture one at a time for best results.
Best practices
- Overhead angle
- Good light (natural is best)
- One plate per photo
- Include a known-size object (a fork) for scale if unusually large
What we don't do
- We don't use your plate photos for model training.
- We don't store the raw photo by default — only the structured analysis. Opt in to retain photos via Settings → Data.
Still stuck? Email support@bragi.fit — we reply within 8 hours.