FoodPercept

Predict how food will be perceived. Design for the people who eat it.

We study how the look and make-up of food relate to what people expect before they taste and what they experience when they do. The research is under way; the models built from it are not yet validated. The goal is a perception engine that food teams can use.

Talk to us See the research

Sweetness, rated 1 to 7

Expectedfrom looking
Experiencedwhen tasting
Illustration. Sensory studies ask people to rate qualities such as sweetness from low to high. What people expect and what they taste can differ. That gap is what we study.

The question we study

Can we computationally model how the physical properties and appearance of food shape what people expect to perceive, what they actually experience, and how that perception changes when the food changes?

  • T1 · Expectation

    What do people expect?

    What taste, smell, texture and sound do people expect from how a food looks?

  • T2 · Experience

    What do people actually perceive?

    How do formulation, process and physical measurements relate to what people report when they taste?

  • T3 · Gap

    Where do the two differ?

    When expectation and experience diverge, by how much, and why?

  • T4 · Design

    What change gets us there?

    Which controlled changes to a food move perception toward a target?

What we are building

Product vision

FoodPercept Engine: a perception model that food R&D teams can plug into the tools they already use. It would estimate how a product is likely to be expected and experienced, with confidence measured against how much real tasting panels agree.

Not available yet. We are looking for R&D teams who want to shape it.

Potential applications

  • Taste what mattersChoose which candidates are worth sending to a tasting panel.
  • Check the promiseCompare what packaging and photos suggest with how the product tastes.
  • Reformulate with careExplore less sugar or salt in foods people still enjoy.

How we work

  • Human agreement is the yardstick. People disagree about taste. We measure how well people agree with each other, then hold models to that level.
  • Plans come before results. Each study's test set, metrics and decision rules are fixed before any model is scored.
  • Independent critique. Study plans and claims are challenged by a reviewer who did not run the work.
  • Sources are verified. We cite only work we have checked against the original.

Where we are

Now
Our first study asks how well today's AI models predict sensory expectations from food images, and what drives their predictions beyond knowing the dish. It uses the public FoodSense dataset (Ishraq et al., CVPR Workshops 2026). We will share results after analysis and review.
Next
Small, controlled tasting studies that record expectation before tasting and experience after it, on the same foods, subject to ethics review.

Who we are

FoodPercept is led by Tony (Trung Tran Thanh), an AI researcher and engineer based in Hanoi, working with a small AI-assisted research team. We want to hear from food scientists, sensory researchers and R&D teams.

[email protected] Talk to us