DubHacks '24 Hackathon

LGBTQ+ Mental Health
AI Assistant

Hilights

At DubHacks 2024, I led UX design and AI prompt engineering in a 24-hour project that won 2nd place out of 69 teams in Intel’s AI Challenge. Our AI-powered mental health assistant provides accessible, empathetic support specifically designed for LGBTQ+ youth, understanding their unique challenges and needs.

Time: October 2024 (24 hours)
Role: Lead UX Designer & AI Prompt Engineer
Involvement: AI persona development, prompt engineering, UX/UI design, frontend development
Team: Collaborated with 2 engineers

tl;dr

In Summary, here's what I accomplished:

  • Led rapid prototype with AI development approach, focusing on core functionality before UI
  • Developed an AI therapist persona with LGBTQ+-specific knowledge and empathetic responses
  • Fine-tuned AI model responses through 2 rounds of user testing with LGBTQ+ participants

🥈 The project won 2nd place out of 69 teams in the Intel AI Challenge

  • Completed highest number of functional features in 24 hours
  • Received mentor recognition for potential partnership with Amazon’s Glamazon

Problem

LGBTQ+ Teens Face Mental Health Struggles with Limited Support

Users

  • Approximately 15 million adults identify as LGBT+ in the US, with higher percentages among the younger generation.

Problem

  • LGBTQ+ teens are 6x likelier to have depression symptoms than non-LGBTQ+ peers
  • 57% of LGBT+ individuals have experienced threats or non-sexual harassment.

Poor Current Solution

  • Over 50% of LGBTQ+ survey participants experienced healthcare discrimination, including denial of care, harsh language, or illness blame based on identity.
Participant 4

…if some kids are only 14, they might face issues like not knowing whether they should come out. And they might not have enough money for a therapist.

How Might We

Support LGBTQ+ teens in finding easy access to mental health assistance?

UX Challenges

  • Limited access to affordable, private mental health support
  • Fear of discrimination and lack of LGBTQ+-specific resources

Technical Requirements

  • Build functional Intel AI-powered prototype in 24 hours
  • Ensure secure, real-time response with scalable model

Strategy 1​

Rapid Prototyping with AI-First Development Approach

Built functional chatbot interface in 4 hours by leveraging AI for rapid prototyping and connecting with Intel’s Gemma API – before focusing on UX improvements.

AI-First Approach

  • Used LLM (ChatGPT) to quickly generate chat function interface with Bootstrap components
  • Connected frontend interface with Gemma API before complex UI design

Collaboration with Engineers

  • Paired with backend engineer on Gemma API integration
  • Quick iterations between AI responses and UI

Generate functional interface with ChatGPT

AI-First Design with Chat GPT

Early interface (generic and robotic responses)

Great, we have a working chatbot!
But… it sounds like a polite robot.

“I hear you. Can you tell me more about how you’re feeling?”

Strategy 2

Response Engineering & Model Fine-tuning

Developed AI responses through prompt engineering and fine-tuned model behavior to create empathetic, LGBTQ+-aware mental health support.

AI Behavior Design

  • Engineered prompts for key mental health scenarios
  • Set up crisis detection and support resources

Model Fine-tuning

  • Tested with LGBTQ+- users with real coming-out and breaking up scenarios
  • Noted unclear/inappropriate responses
  • Adjusted tone and responses based on user feedback

Impact

  • Natural, empathetic conversations
  • Appropriate handling of coming-out discussions
Scenario

LGBTQ+- Key Mental Health Scenarios

Designed conversation starters for common LGBTQ+ concerns, allowing AI to provide more targeted and relevant support from the first interaction

Questionnaire

Crisis Detection & Support Resources

  • Built-in PHQ-9 depression screening
  • Crisis triggers for self-harm language
  • Immediate professional help referrals when needed

Model Fine-tuning Based on User Feedback

Before

AI: I understand how you feel. Coming out must be hard.

User Feedback

Participant 4
  • It keeps saying “I understand” but doesn’t show it. It feels weird and robotic.
  • I shared I’m scared to come out, and it just said “that must be hard”. I think it would be better if it ask some follow up questions.

Prompt Enhancement

❇️ Added specific follow-up questions instead of generic acknowledgments
❇️ Trained AI to listen more and offer support before suggestions

After

AI: It takes courage to think about coming out. Many people feel a mix of emotions during this time. What feelings are coming up for you right now?

Strategy 3

Technical Communication & Alignment

Led cross-functional collaboration through flow charts and user stories , turning complex AI features into achievable 24-hour goals.

Communication Challenges

  • Limited time to align on technical scope
  • Different vocabularies between design and engineering (I said “design AI response,” engineers called it “create AI persona”)

My Solutions

  • Created user flow charts mapping user flow & AI decision points
  • Prioritized features based on technical feasibility

Reflection

🥈 2nd Place in the Intel AI Challenge

Hackathon Intel 2nd Place

Team Trust and Alignment

Judges noted our team had the highest completion rate. This was due to our early team alignment and trust, which kept us on track. I loved the strong collaboration and support within the team—it made the process smooth and enjoyable.

hackathon_getting feedback from Amazon mentor

Feedback and Opportunities

An Amazon mentor suggested potential partnerships with groups like Glamazon and medical experts to expand our AI’s impact, especially for users facing severe mental health issues. This feedback opened my eyes to future possibilities for the project.

Final Product

A hardware design leveraging data visualization to transform testing insights into ergonomic monitor solutions.

A food traceability system leveraging AI image recognition to transform production data into streamlined quality control standards.