A smart scale that sees food waste.
For Phood Solutions in the United States we built Phood 2.0: a scale, a camera and a food classifier that turn every item a kitchen throws away into a record.
The problem
A kitchen cannot cut waste it does not measure. Food goes in the bin all day, at every station, and nobody stops to write down what it was, how much it weighed or why it was thrown away.
Phood 2.0 had to capture that record at the moment of disposal, with almost no extra work for the cook holding the pan.
What we built
Three parts, each doing one job.
- An Android app running on an industrial touch scale. It read the weight from the scale over its serial protocol and captured a photo of each item.
- A food classifier: a ResNet-20 model built in Keras, running on AWS Lambda, that named the food in the photo.
- A Node.js and Express API that held the kitchen's world: organisations, locations, stations, pans, food logs, waste reports, loss reasons, menus, inventory and roles. Validation, authentication and caching ran as middleware in front of every route.
How it worked in a kitchen.
Five steps for every item, most of them automatic.
- 1. WeighThe cook sets the pan or item on the scale. The app reads the weight over the serial connection.
- 2. PhotographThe app captures a photo of the item on the scale.
- 3. ClassifyThe photo goes to the ResNet-20 classifier on AWS Lambda, which names the food.
- 4. LogThe API stores a food log with the weight, the food, the station and the loss reason.
- 5. ReportLogs roll up into waste reports for each location, ready for the kitchen manager.
What we'd do today
This is our opinion, looking back. In 2017 a small custom classifier in the cloud was the practical choice. Today we would try two things first.
We would run a compact model on the device itself, so the scale keeps working when the kitchen's network does not, and a photo never has to leave the building unless it needs to.
For the hard cases we would test a modern vision model before training our own. It can name a dish it has never seen and describe what is in a mixed pan, which a fixed list of classes cannot.
What we would keep is the shape of the system: one action at the bin, everything else automatic, and reports a kitchen manager can act on.
Phood has since grown its computer-vision platform into PHOOD X and, by its own figures, has analysed 1.8 billion pounds of food. phood.com
Tell us what you're trying to solve.
A 30-minute conversation with Nader. You leave with a clear next step, whether or not we work together.