Product transparency

Nutrition estimate methodology and limitations

Keola is designed to make consistent meal logging easier. Its output is an informed estimate, not a direct measurement of the food.

How an estimate is produced

  1. The user sends a meal photograph, a written description, or both through WhatsApp.
  2. A multimodal AI model identifies likely foods, visible portions, and preparation clues.
  3. The model estimates calories, protein, carbohydrate, fat, and fibre using typical reference portions.
  4. Keola checks and totals the structured result, then returns it in the conversation.
  5. The user can correct portions or ingredients conversationally. User-supplied details take priority over visual inference.

What the system cannot observe reliably

A photograph rarely reveals exact weight, hidden cooking fats, every recipe ingredient, or how much was consumed. Containers and camera angle can distort scale. Mixed dishes are especially uncertain because calorie-dense ingredients may not be visible.

How to get a better estimate

  • Show the entire plate in clear light.
  • Include a familiar object or state the portion size.
  • Name hidden oils, sauces, dressings, and calorie-dense additions.
  • Say whether the shown serving was fully eaten.
  • Correct the response when Keola identifies an ingredient or amount incorrectly.

When not to rely on a photo estimate

Use weighed ingredients, verified packaging data, or qualified clinical advice when precise intake is medically important. Keola does not diagnose, treat, or replace a healthcare professional.

How performance is tested

The public benchmark runs the same food-image analysis used by Keola against a fixed sample from Google Research's independently labelled Nutrition5k dataset. It reports aggregate error and individual results, including failures. This is a limited test of cafeteria-style meals, not a guarantee for every cuisine, camera, or portion.

View the benchmark and reproduction details.

Change log

6 September 2026: First public methodology version. Added the Nutrition5k benchmark protocol and explicit clinical-use limitations.