Abstract:
Keeping a balanced and individualized diet
plays an important role in the prevention of chronic diseases
as well as in maintaining optimal wellness. The accurate
monitoring of daily calorie consumption and abiding by the
restrictions of certain diets is key to achieving good health.
Nevertheless, existing dieting apps are not dynamic enough;
they only serve as online diaries incapable of adjusting
according to the behaviour of the user. A multi-model
machine learning architecture was developed for the
MyAiPlate application to address this gap. A YOLOv8 object
detection model, fine-tuned on a real-world dataset of 1,169
labelled food images, was implemented to visually recognize
food items and estimate caloric intake from user-captured
images. The proposed machine learning components were
rigorously evaluated for performance and user safety. The
YOLOv8 food recognition model was validated using
unseen, real-world meal images. On the held-out validation
set the model achieved 95.16% precision, 94.42% recall,
97.51% mAP@50, and 75.57% mAP@50-95. In addition to
encouraging user participation, feedback was provided
through a conversational interface within the application.
Traditional diet tracking tools do not have a conversational
approach and cannot answer specific user inquiries in a
context-aware manner. An AI Nutrition Chatbot was
implemented using Large Language Models (LLMs) based
on transformers from OpenRouter, which offered real-time,
context-aware, personalized nutritional advice based on meal
entries. The conversational interface was tested using 50
queries, including 25 challenging nutritional queries and 25
common diet prompts to ensure that the LLM could deliver
reliable results without hallucinations.
