Mendy
An Empathetic AI
An Empathetic AI
Our
Team
Jaylen Luc
Jibreel Rasheed
Matthew Cho
Neal Lowry
Shiyi Mu
Our Sponsor
Dr. Adam Starks
A motivational speaker on child welfare, youth homelessness, and juvenile justice, Dr. Adam Starks fights to mitigate the struggles of traumatized youth. Experiencing firsthand the adverse effects of the foster care system, Dr. Starks inspired the creation of Mendy to combat the struggles of at-risk youth through research-based empowerment.
The Problem
Although talking about mental health has been largely de-stigmatized, starting the conversation is still especially intimidating and inaccessible to many adolescents due their lack of information and resources. Fear of judgment, being misunderstood, and feeling like a burden to others all deter individuals grappling with mental health issues from seeking the help they need.
The Solution
Our solution to this problem is Mendy, a conversational AI that creates a low-stakes, non-judgmental space for those facing mental health challenges to express their struggles and receive personalized validation. Even though our target audience is adolescents, Mendy's value extends well beyond this group. Mental health struggles are universal and indiscriminate, thus Mendy was built to be capable of offering support to anyone facing mental health challenges, regardless of age.
Before beginning development we established a consistent schedule for progress updates with our sponsor, which proved invaluable for maintaining team accountability and adherence to Dr. Starks’ goals and expectations.Throughout each development cycle the trajectory of our system changed significantly, and these regular meetings enabled our team to stay on the same page and make continual progress.
The most important task during our project cycle was finding the right tech stack. Substantial effort went into determining the right tools as choosing incorrectly could potentially require re-engineering everything from scratch well into development. The system architecture was then blueprinted to ensure all variables were accounted for, followed by designing, with coding coming last due to its time intensive nature. Development was finally executed on a feature by feature basis to promote measurable progress and mitigate dependency issues come deployment.
A document encapsulating the diverse needs of Mendy's stakeholders, showing the wide range of utility our minimum viable product would need to have. Each story highlights how Mendy should support its users, guiding our development process by informing us on what features are crucial to optimally support our users given the sensitive context.
why created, how helped, what it shows
Personas representing in detail the context of those who would benefit from mental health support. This document helped narrow our intended audience to those who are without resources and information, as individuals who don't have access to professional mental healthcare resources could potentially utilize Mendy as a placeholder alternative.
A storyboard showing the journey of a user engaging with Mendy to navigate challenges related to homelessness and hunger. This graphic helped us ensure the user flow was intuitive and efficient as those in need of Mendy's reassurance may not have extensive time to spend online and would benefit from a streamlined system.
Since the inception of the idea of an empathetic chatbot, Dr. Starks had grand ideas on how to navigate the nascent field of artificial intelligence to support struggling youth. This led to the unforeseen need to significantly refine project scope, as within the extensive features originally planned there were many unnecessary and circumstantially detrimental features that impaired the user experience. For instance, originally it was planned to split the site into two sections, one where users in need could search for local resources such as food and shelter, and the other for the empathetic chatbot. The ability to search was ultimately scrapped to make the project feasible as it was decided that Mendy could provide resource direction as necessary given the context of the conversation.
Another significant challenge encountered was the ethical concerns raised by artificial intelligence in a healthcare setting. While AI has the potential to revolutionize treatment plans, its generative nature can yield inaccurate or insensitive results. AI’s reliability cannot yet be trusted without supervision, making its integration into the healthcare sector potentially detrimental. To combat this, our site includes disclaimer stating that Mendy may provide inaccurate information. We also made sure to meticulously train our LLM using over 76,000 conversations to maximize the probability of providing appropriate responses, especially given sensitive queries pertaining to potentially dangerous or harmful situations.
Since our initial prototype, substantial improvements in Mendy’s responses were made, largely based on insights we gained through user testing and building the system itself. An originally unforeseen problem user feedback uncovered was that Mendy had a tendency to be excessively verbose. In one case, a user entered the query “I love lemons” and Mendy provided a three paragraph response about lemons, thus responses were made more concise when possible to enhance user experience.
Another issue arose when we began to implement our mockups as we realized the fun, bubbly aesthetic we had designed seemed insensitive given our mission to support individuals who feel overwhelmed and might prefer a safe, relaxing environment. We redesigned our site to be simple and calming, incorporating a slowly breathing gradient background to ensure our users felt comfortable and supported.
Mendy is built using Next.js for frontend development and Django for backend development. Vercel and Heroku are used for app deployment. The core model is trained using the Stanford Emotional Narratives dataset and leverages AstraDB as a vector store for retrieval-augmented generation. Firebase is used to preserve chat history, providing efficient and scalable storage. User authentication is managed using JWT.
In addition to code files and appropriate credentials to the respective accounts, further deliverables include the following: a revised ethics review addressing concerns in our software, an architecture diagram presenting the interactions between core features, and an R&D document detailing our iterative design process through various prototypes.