Nutritional AI
AI & Data
Semester programme:Minor AI for Society
Client company:Just Eat Takeaway
Alaa
Baris
Zinho
Tamas
Laurens
Project description
How can we create a food chatbot that will make proper, well informed food recommendations to its users?
Context
This project is in the food delivery and health domain. It takes place at Just Eat Takeaway, a food delivery platform. The project focuses on helping users make healthier food choices when ordering meals online. The team is building an AI chatbot that suggests healthier meals in a friendly way, without limiting the user’s freedom to choose what they want.
Results
The main outcome of this project is a working AI-powered chatbot that suggests recipes based on available ingredients, number of servings, and a specified budget. Built using Python, LangChain, and a local large language model via Ollama, the chatbot functions within a Streamlit interface designed to mimic a mobile app experience. Users can input their ingredients and preferences in natural language, and the chatbot returns feasible, budget-conscious recipe suggestions, including what still needs to be bought and estimated costs.
This product was tested in a local environment and validated through practical user trials with a small group of seven test users. Of these, five indicated that the chatbot’s suggestions were usable, clear, and relevant. These results suggest a Technology Readiness Level (TRL) of 4, meaning the technology has been validated in a controlled setting. While not yet deployed in a real-world production environment, the prototype reliably delivers its intended function and lays a solid foundation for further development.
In addition to the functional chatbot, a minimum viable product (MVP) of the visual mobile interface was created. This design was guided by principles from marketing psychology, specifically nudging theory and decision architecture. The visual version was tested through A/B comparisons using AI-generated mock-ups. In a preference test, 71% of participants favored the version with simplified recipe suggestions and appealing food visuals. These early design tests confirm the potential value of visual nudges in promoting healthier or more budget-conscious choices, corresponding with findings from literature on visual attention and cognitive load. This part of the project is best positioned at TRL 3, as it represents a conceptual validation with initial user feedback.
Throughout the process, several insights were gathered that are relevant for future development. First, simplicity consistently outperformed detail: users preferred brief, focused recipe suggestions over long-winded explanations. Second, the chatbot performed surprisingly well with vague or incomplete user input, as long as prompt engineering was carefully adjusted to support flexible reasoning. Third, applying visual nudges — such as highlighting cheaper or healthier options with color or layout — significantly improved perceived usability and choice satisfaction. These insights are backed both by small-scale validation and by academic literature (e.g., Thaler & Sunstein, 2008; Wansink, 2016), and they offer valuable direction for moving toward TRL 5 or 6, where real-world deployment and iterative user testing would take place.
In summary, the project delivered a functional, validated prototype, a first-stage visual design grounded in behavioral science, and multiple validated insights into user interaction with AI in a food-choice context. Together, these outcomes form a strong base for future refinement and scaling.
Download Research Document (PDF)
About the project group
Except for Laurens all group members have an ICT background. Laurens studies Applied Psychology and Baris also has a background in the education Game Development. The project group has a great spread in different ICT subjects.