ASSIST Lunch Management

ASSIST Lunch App is an R&D project developed by ASSIST Software, using Django. Its main purpose is to allow employees to order lunch from a list of restaurants that we have a delivery contract with. Every day, restaurants post their menus, and colleagues order based on their preferences. The app also generates various monthly reports to help with accounting.   

Auto Order is a submodule of the Lunch App. It automatically orders relevant dishes for users when the option is enabled. The module uses order history to train the models for each user individually. We’ve developed our own Neural Network, with a simple implementation that does not rely on external ML libraries. 

We are using a shallow network with a single hidden layer of 8 neurons. Tests with more hidden layers did not provide a significant accuracy increase and were slow performing. We’re working on new updates to attain higher accuracy. Currently, a user model is trained for a maximum of 30 minutes or until it reaches 80% accuracy.

Since there are many unique dishes (around 3000) we could not use a classification approach, so instead we used an NLP/NLU strategy. The vocabulary consists of 250 unique words that make 96% of all words and are used at least 10 times in the dish names listed by the restaurants since 2013. There were a little over 500 unique words used for the dish names.  

We’ve also used as input the day the order was made, the month, the number of days passed, the restaurant, and the type of food: first/ second course or fitness meal. These parameters help detect any periodical patterns that a user might have in their food preferences. Also, a dish might be prepared in different ways at restaurants. This is a case of Supervised Learning. 

Features we developed:   

• Front-end 
• Back-end 
• UI/UX 
• Supervised Machine learning algorithms 
• Office 365 Sync 
• Text & Category interpretation 
• NLP and NLU 
• Monthly reports for accounting