Preparing your next favourite dish with the help of Matprat's chatbot

Photo of fajitas with side dishes for the MatPrat project.

Preparing your next favourite dish with the help of Matprat's chatbot

MatPrat’s recipes and inspirational content, with detailed instructions, are now available through a live chatbot.

Deliveries

Completed processes for user insight and user testing
Designed user dialogue and functionality
Developed Norwegian language model and modules for support vector machines
Developed a chat-bot able to run on Facebook Messenger
Developed a tailored administrative module for editing and training the chat-bot
Implemented Elastic Search as the search engine for content in EPiServer and Youtube
Configure Kibana to show statistics

Technologies

Elastic Search
Rasa NLU and Rasa Core
Docker
Microsoft Bot Framework
Azure
Kibana
Kubernetes

Team

Andreas Schafferer
Frode Rennemo
Ida Noemi
Thomas Walter van Deurs
Eirik Stavelin
Bjørnar Remme
Christine Raybo
Terje Nygård
Pål Rønning

“I have never worked on such an advanced and complex project - both technically and design wise. This has truly been fun!” Thomas Walter van Deurs, Senior Delivery Manager, Making Waves Oslo

“I have never worked on such an advanced and complex project.”

Challenge

The starting point for the service was to reach users on a new platform with some sort of user interaction and a digital interface through a chatbot. The service was to match content from MatPrat with the ingredients you’ve got on your hands, regardless of them being whatever is hidden in the back of your fridge or the result of shopping on an empty stomach. The goal was also to provide search results for specific recipes and general cooking advice. The service was to use machine learning in order to provide as relevant search results as possible and base itself on its very own Norwegian language model, specifically aimed at food.

A bowl with sweet potatoes, avocado, salad and pomegranate on a table.

Our approach

Working interactively, the project started out as an early technical prototype, to prove that the chosen technical components complement each other in the expected way. Creating a Norwegian language model was considered the most complex element, and this work was started early. We also began an insight stage to get an overview of the users’ needs, premises for the service and to decide which specific concept the service should address. Different samples of the chatbot were defined and regularly presented for the users to try - so our recipe was adjusted along the way.

Matprat's green and offwhite chatbot, Mats, shown floating in a blue room.

The outcome

The service is just out of the oven, but has so far gotten a good reception of more than 650 users on its first day. Users are commenting how well it works and that it presents MatPrat as an innovative stakeholder.

You can connect with Mats here.