Novu.
A calmer way to ask,
"am I okay?"

To evaluate our design, we conducted two rounds of usability testing with two different groups of students (Figure 2.1). Each session asked users to perform a set of tasks from our low-fidelity prototypes (Figure 2.2), including navigating the main menu, accessing the symptom based assessment mode and the mental health mode, and other additional features like the add file/image buttons and finding their way back to the menu from a conversation. The goal was to observe how naturally users moved through the app, identify points of confusion, and understand how they felt about interacting with an AI in a health context.

Both rounds of testing confirmed that the interface was simple and easy to navigate. Users in the first session were able to access the different modes without much friction and understood the general purpose of the app quickly. By the second session, users were navigating the application and switching between modes intuitively, which suggested that the core layout was clear and learnable.

User Test 1: These group members navigated switching between different modes and explored the features of our interface. The application is designed to give either physical health advice or mental health advice. This group gave suggestions for interface and user experience improvements, and asked questions on how the chatbot would respond to irrelevant questions (e.g. if it wasn't related to health in any way). It was something we hadn't thought about, so that was something we considered when thinking about how to prompt our AI. They also helped us discover the need for a back button.

User Test 2: These group members were asked to navigate the main menu and additional buttons, and share their thoughts on the app's features and which mode they would be more inclined to trust and use. Their feedback surfaced two key feature recommendations: a memory function that allows NOVU to reference past conversations to provide more contextual responses, with the option to clear that memory for privacy, and a voice chat feature for both the user and the chatbot to make interactions feel more natural, particularly within the mental health mode.

Overall, we learned that users were more comfortable receiving symptom-based guidance for physical health concerns. This distinction directly shaped how we approached each mode's tone and design. Users also raised concerns about data privacy and how their profile information would be used to personalize responses, signaling that trust and transparency need to be built into the experience, not just assumed. We also learned that if the chatbot identifies a potential medical emergency, users expect a clear recommendation to contact emergency services, which we incorporated as a design requirement. Across both rounds of testing, the clearest takeaway was that our core design decisions were sound. Users found the interface intuitive and were able to navigate it without significant guidance. Where the testing proved most valuable was in surfacing gaps we hadn't anticipated: how the chatbot should handle off-topic inputs, the need for basic navigation controls like a back button, and the importance of giving users control over their data through features like memory clearing. (Figure 2.3)

In addition to usability testing, we conducted a heuristic evaluation of the interface using Nielsen's 10 usability heuristics (Figure 2.4). Overall, the design held up well. The minimalist layout reduced cognitive load, features were clearly surfaced without requiring users to recall them, and the profile-based personalization supported flexibility and efficiency. The most notable gap identified was around user control and freedom: without a stop-response feature, users who made a typo or wanted to rephrase a question had no way to interrupt the AI mid-response. We also flagged that hallucination risk in a health context requires stringent safeguards, since the consequences of inaccurate information are higher stakes than in a general-purpose chatbot. These findings informed our next iteration priorities.

The problem our team chose to address was: how do we get patients who are losing quick access to medical services to trust and use an AI insight tool to ease their concerns and get quick answers/possible solutions to health anxieties they might be having? When looking at the problem and our design, it feels as though we were able to do that successfully. The key thing was making something accessible to people. Our design is available to everyone, and it includes accessibility features such as the Voice-to-Text feature (Figure 4.3g) that make it even more accessible. As well as in the settings menu (Figure 4.3d), where users can make further adjustments to Novu and personalize it to make it easier for them to use. The second reason it addresses the problem is that we took into account all the feedback we received from user research and usability testing. Something we kept hearing was that people were concerned about their privacy and data. When we look at the settings menu (Figure 4.3d), we see that the user has access to the security and data controls tabs, allowing them to make adjustments as they please over time. As well as clearing the memory, so your previous conversations and the personal context the bot learns about you are erased. Adding features like the Symptom Assessment mode (Figure 4.3f) and the Mental Health Support mode (Figure 4.3e) allows users to use it for various reasons, making them more likely to trust Novu to ease their health anxieties. Novu is capable of answering medical questions, providing medical advice, conducting symptom assessments, and offering mental health support. It is available 24/7 and is a place where people who have lost quick access to medical services can go to ask questions about health concerns.