NUTRI AI KITCHEN SCALE

Overview

Nutri is an AI-powered scale that recognizes food as it is being weighed, helping users easily track nutrition and macros in the kitchen. Built on TinyML developed at MIT, Nutri runs food-recognition models locally on low-power hardware without relying on the cloud or a phone. I helped design the scale, embedded interface, and companion app as one integrated system, which went on to receive a 2022 Red Dot Award for Design Concept.

The things that made this project difficult:

AI had not been popularized yet, so the idea of bringing a machine learning experience into your home was foreign to most people.
A kitchen device has to earn its counter space or it ends up lost somewhere in a drawer.

Nutri demo video

The problem

The standard workflow needs a scale, a nutrition label, a calculator, and a notebook. With Nutri you turn on the scale, cook, and the tracking happens on its own.

Current user journey

Five line drawings in a row: gather a scale, calculator, notebook and nutrition label; weigh the food; find its nutrition facts; calculate the serving; write each value down

Ideal user journey

Three line drawings in a row: turn on your scale; the camera recognizes food as you weigh it; the nutrition value is recorded automatically on the scale and the phone

Research

We interviewed four primary users and pulled from secondary sources to map out how people track their food. We built two personas, a Type 1 diabetic and a gym enthusiast. We chose them because both groups already track their diet closely and already own food scales, and a new product is an easier sell when it fixes a process people are stuck doing anyway. Three insights came out of the research.

The scale has to work alone for simple tasks. Sometimes you just want a weight, no sign-in, no phone.

The hardware and software experience together is the reason to buy. Cooking cannot slow down for tracking.

Users give a lot of data and get little back. The app has to return more than it takes.

Persona sheet for Sarah, 24, a graphic designer with Type 1 diabetes: goals, concerns and a photo of her weekly meal prep

Persona sheet for Josh, 33, a project manager and gym enthusiast: goals, concerns and a photo of his meal prep containers

The market

Smart food recognition already exists in industrial settings, but there are no consumer products tackling this space.

Two-by-two market map, simple to smart across and consumer to industrial down: consumer food scales cluster at the simple end, industrial recognition systems at the smart end, and the smart-consumer quadrant is marked as the gap

Tiny ML

Running the AI on the device instead of the cloud or the phone made the scale cheap, private, and self-sufficient.

Scenario mapping

Before designing the hardware or UI, we mapped out some of the most common cooking scenarios a scale runs into: from-scratch baking, half-scratch pasta, meal prep, premade food. Each scenario got its own screen flow, and those flows drove both the interface and the hardware.

Hand-drawn screen flows for six cooking scenarios, one per row: baking from scratch, pasta from scratch, whole ingredients, meal prep, meal prep with protein, and premade food, each a chain of small scale sketches with notes

The scale

The camera arm folds flat so the scale stores like a cutting board. The camera angle is fixed so the shot is right every time. A light shows when food is in view, and the glass top wipes clean.

Two white Nutri scales on a grey counter: one open with the camera arm raised and a bowl of berries on its glass top, one folded flat in front with the arm lying across the base — Open and folded

The same two scales seen from above, halved tomatoes on the open scale beside its screen and the folded scale below it — Open and folded

Prototypes

3D printed prototype parts of the scale which we used for user testing and for filming a usage video.

Flat lay of 3D printed scale parts on a white ground: the top shell with its screen cutout and UNIT and power/tare marks, the bottom shell with ribs and screw bosses, a printed camera arm with its pin, a flat base plate, a grey plate with a window for the load cell, and three arm sections

A person taking the folded scale out of an upper kitchen cabinet, next to a stack of plates — Easy to store

A person chopping peppers at a kitchen counter with a salmon fillet on the scale beside the cutting board — Easy to store

What lives where

The on-scale screen does weighing, recognition, confirm, and save, and the optional app does the rest. The one overlap is Live Scale, a real-time view of the scale inside the app for when you're multitasking or a large pot blocks the screen.

Designing the app

Major features of the app include dashboard, meal logging, recommendations, and personal assistant.

Information architecture

Pen sketches of phone screens: two versions of the add-meal flow from meal type to food item, four sign-in page options, a health test flow, and two dashboard layouts with calorie and macro rings — Screen sketches

Mid fidelity prototype

Onboarding question 3 of 5: what are your primary nutrition goals, with weight management and dietary pattern options — Sign up

Feedback and next steps

I user tested the high-fidelity version, and the feedback fell into three categories. The project wrapped before we could act on these, so they stand as the direction a next version would take.

Data visualization

The circular charts were hard to read and progress was not glanceable.

Hierarchy

Too many accent colors competing, so the most important number did not stand out.

Personalization

Users need to get more than they give.

Red Dot Boards

In the end, tracking a meal went from five steps to almost none. You just cook and the scale does the rest. This project was a recipient of the 2022 Red Dot Award in the concept category.

Red Dot submission board 1 of 5: Nutri, nutrition tracking simplified, with the hero render and the problem, technology and opportunity — The five boards submitted to Red Dot

Board 2 of 5: our audience and the standard five-step food scale workflow against the three-step Nutri workflow — The five boards submitted to Red Dot

Main takeaways

A collection of thoughts and challenges throughout the project.

AI ethics

In health products, AI suggestions need real testing before release, and users should still be able to think without the AI, the way you should still be able to drive even if you own an autonomous car.

Data translation

Information hierarchy and visual hierarchy decide whether people understand their own data.

Test early and often

Low-fidelity digital prototypes were the fastest way to work out how the hardware and software fit together.

Two white Nutri scales on white: one open with the camera arm raised, halved tomatoes on its glass top and the screen reading 114.23 g, one folded flat beside it