Throughout June, I participated in the Edge AI Summer Program exploring the exciting potential of Edge AI in medical devices. We investigated how artificial intelligence (AI) can be embedded within devices to analyze data and provide real-time feedback for applications like diabetes monitoring devices. The program culminated in a collaborative project where we designed and developed an AI-powered solution. This teamwork experience allowed me to not only gain a deeper understanding of Edge AI concepts but also apply them to a real-world scenario alongside a talented team. I’ll cover the basics of our project in a later section.
Remote Experience
The Edge AI Summer Program is a collaborative effort by the National Science Foundation (NSF) with universities across North Dakota and Alabama specifically designed for undergraduate students. The first two weeks of the program were fully remote. This provided a foundation in key areas of Artificial Intelligence (AI), including the basic network structure of models and the differences between models. This foundational series of presentations allowed us to build a strong foundation in AI fundamentals, especially for those new to the field.
Once all the undergraduate participants had a grasp on the basics of AI, professors and PhD students at the forefront of AI development showcased their cutting-edge work, giving us a glimpse into the exciting possibilities of the field. These works focused on how AI can be applied to the medical field. These technologies included designing models to diagnose disease based on medical imaging, creating AI-integrated diabetes monitoring devices, an Alzheimer app that helps patients keep their minds active and track mental decline, and many other similar technologies.
Many of these technologies shared with us focus on one day integrating their AI models into edge devices. These edge devices, essentially any device at the network's 'outward edge' that collects or communicates data, can range from simple sensors to complex medical equipment. Examples of edge devices include your phone, computer, or smart TV. The goal is to enable Edge AI, a field that equips these devices with built-in processing capabilities to analyze data and generate insights directly at the source, rather than relying solely on sending data to the cloud for processing. For instance, when you use AI-powered services like ChatGPT, the complex AI model doesn't run on your phone, but on a powerful remote server. These researchers aim to develop AI models that can run on these less powerful devices with similar proficiency.
 |
|---|
| Diagram of the network of cloud and edge devices |
On Campus Experience
 |
|---|
| Campus of North Dakota State University in Fargo, North Dakota |
The remote learning portion provided a solid foundation, but the real magic happened during the in-person residency at North Dakota State University in Fargo, North Dakota. Undergraduate researchers from across the country converged for an intensive two-week experience. The first week delved deeper into the research we'd encountered online. We weren't just passive listeners anymore! We had the incredible opportunity to tour the labs of researchers at NDSU. Witnessing firsthand the researchers' dedication and the sophisticated equipment used to push the boundaries of human knowledge was truly inspiring.
 |
|---|
| On a side note, we got to see how an electron microscope works in one of the labs. It’s crazy that something can be viewed in up to 10,000,000x magnification. This is what a butterfly wing looks under one! |
Final Product
The focus of the final week shifted to our own research aspirations. As groups, we explored areas of AI that interested us. This hands-on experience allowed us to not only apply the concepts we learned but also gain valuable teamwork skills in a fast-paced environment. My group focused on incorporating a large language model (LLM) in a medical setting. We know access to mental health care is critical to a patient’s overall well-being. However, some individuals feel uncomfortable sharing their struggles, others find it expensive, and many do not have the time to seek help. Noticing these problems we utilized Google’s Gemini LLM to perform the actions of a therapist and assist individuals experiencing depression by providing empathetic and supportive interactions.
 |
|---|
| Therapy Bot's Homepage |
 |
|---|
| Therapy Bot's Chat Feature |
Our final project Therapist Bot can be found at https://therapybot.netlify.app/.
Exploring the Technologies in Therapist Bot
The core functionality of the chatbot was powered by Google’s Gemini LLM provided by Gemini’s API service. The service provides advanced natural language processing capabilities. Google’s LLM was trained on a large dataset, making it versatile to many prompts or instructions. To ensure the chatbot could effectively assist users experiencing depression, we prompted the model to consider this prompt before every response, “You are a therapist chatbot designed to help depression patients cope with their symptoms. Comfort them and make them feel heard. Remind them, you are not a licensed therapist so make sure to recommend them to further medical resources if needed”. This guiding prompt enabled the chatbot to provide empathetic and supportive responses while maintaining a clear boundary that it is not a licensed therapist.
The application’s front end was designed using the Svelte framework, Skeleton UI Toolkit, and Tailwind CSS. Svelte was chosen for its efficiency and performance benefits, enabling us to write concise code while achieving high reactivity and interactivity. Skeleton UI Toolkit provides premade Svelte components with Tailwind styling, it allows for fast frontend development. Tailwind CSS, a utility-first CSS framework, allowed for rapid development with its pre-defined classes, ensuring a consistent and responsive design.
The back end of the application was designed using SvelteKit, a backend framework built on Svelte that allows for seamless integration. The backend server takes any patient’s response and fetches a response from Google-Gemini’s API. Gemini’s response is tailored to a patient’s current conversation history and is based on the guiding prompt mentioned previously.
Results
Our therapy-focused chatbot displayed that it could provide basic therapeutic advice to patients. The chatbot not only displayed competency in offering basic therapeutic techniques, but it also demonstrated a remarkable ability to provide emotional support. The interactivity of the chatbot was particularly impressive. The dialogue flowed seamlessly, mimicking a natural human conversation so convincingly that some uninformed users might have mistaken it for a real person altogether. This lifelike quality has the potential to be a significant asset, especially for individuals who may feel apprehensive about seeking traditional therapy.
Experience within the Program
I am incredibly grateful to have participated in this summer program. I had the opportunity to work with so many researchers and undergraduate students passionate about the same emerging technologies as I am. Through hands-on projects and insightful discussions, I not only gained a deeper understanding of Edge AI but also learned valuable collaboration skills that I can apply to future endeavors.
