From photo to macros: how Lorie's AI works

Photo-to-calories sounds like magic — but there's a very deliberate architecture underneath. Not a model guessing numbers; a model identifying ingredients, then handing the math off to data.

Two separate layers#

Lorie's vision model is a fine-tuned multi-label classifier paired with a portion-estimation head that uses contextual cues — plate size, hand for scale, depth.

We don't ask the model to guess macros. We ask it to identify ingredients and estimate grams. Then we cross-reference against TKPI data to compute calories and macros.

Why keep them separate?#

This separation is on purpose. It keeps the model honest, keeps our nutrition data auditable, and lets us update either layer independently — for example, updating TKPI entries without retraining the vision model.

The result: when you photograph nasi goreng, Lorie doesn't "guess" 350 kcal. It identifies rice, egg, oil, kecap — then sums macros from each ingredient based on estimated grams.