SmartFood : personalized food and restaurant recommendation app using LLM

Abstract

This thesis addresses the development and evaluation of a mobile application designed to give users personalized food recommendations based on their preferences using LLMs. The application was developed using Flutter, an open-source framework for building multi-platform applications. Firebase was used for user authentication, feedback collection.

Database is based on the Fineli Open Data package provided by the Finnish Institute for Health and Welfare. The data was used for calorie estimation and display them to the users. The Fineli database served as a reliable starting point thanks to its detailed nutritional information, but notable challenges were identified, for example missing data on certain dishes or portion sizes.

Functional testing confirmed the application's core features to function correctly, including menu scraping, feedback collection, authentication and calorie estimation. While most components performed reliably, minor formatting issues in LLM outputs and varying accuracy of calorie estimations were observed, with single-ingredient items showing consistent results compared to less reliable LLM estimates for restaurant-specific dishes. Performance testing across three Android devices showed acceptable performance. UI responsiveness testing revealed minor overflow issues with longer dish names on certain screen sizes. User testing confirmed the interface to be intuitive and easy to use, but there was some concerns about age information collection.

At the end of the thesis, there are suggestions for future work, which include enhancing the calorie estimation systems, user feedback mechanisms could be implemented to collect real-world data on portion sizes and meal compositions, potentially including photo submissions to improve estimation accuracy. Additionally, the application's dependence on a single data source for menus present a vulnerability.