Why HealthAware talks about foods, not grams

When you photograph a meal, an AI model has two jobs: work out what is on the plate, and guess how much of it there is.

It is very good at the first job. It is much worse at the second. In our own tests, photo-based estimates of nutrients missed by 40% to 77% on average. A bowl of pasta photographed from above looks almost the same at 250 grams as at 350.

The problem with precise-looking numbers

Most food apps hide that uncertainty. They show “1,847 mg sodium” with no error bar, and you start making decisions on a number that could easily be off by half.

We think that’s the wrong trade. A heart patient deciding what to eat tomorrow doesn’t need a falsely precise number. They need to know which food pushed the day off track — because that is the thing they can change.

What we do instead

• We name the foods. Every meal is broken into its foods, and each food is matched to the USDA FoodData Central database.
• We check each food against your plan. Processed meat, full-fat cheese, refined grains — each plan has its own rules, and HealthAware tells you which food broke which rule.
• We suggest a swap, not a cut. “Roast chicken instead of deli turkey keeps the sandwich and most of the salt goes away.”
• Numbers stay where they belong. The meal page still shows the full breakdown with grams if you want it. But the advice — on Home, from Maya, in your weekly notes — talks about foods.

Patterns, not single meals

One salty dinner doesn’t matter much. Cheese on five of the last seven days does. HealthAware looks across days for foods that keep showing up, and only says something when it has seen enough days to be sure. A day you didn’t log is never counted as a day you ate well.

The error is in the quantity, not the identity. So that’s where we put the advice.

If you have feedback on how this feels in the app, we’d love to hear it.