ChatBuddy
Google Gemini AI Hackathon
Designed an AI therapist providing seamless mental health support with personalized, diverse personalities.
Timeline
April- May,
4 weeks in 2024
Team
Data scientist,
Software engineers,
Product designer(me),
Product manager
Role
Information architecture,
Wireframe, User flow,
Prompt engineering
Interaction design
Overview
Context
With recent layoffs, I’ve seen friends and family struggle with work and school pressures. Many want therapy but are put off by long waitlists, high costs, and rigid schedules. This led us to wonder: why not create a virtual therapist using AI for more accessible and affordable mental health care?
Solution
Coincidentally, some friends had the same idea, and we quickly teamed up at the Gemini Hackathon. We aimed to create an AI healing therapist with three distinct personalities using prompt engineering. This AI would offer personalized support, allowing users to choose the personality they need, making mental health care more relatable and accessible.
Mission
We aspire to make mental wellness universally accessible through AI.
Pitch demo (1min)
key features
1/Account login: prioritizing user data protection"
Support only common accounts, not email logins, to protect user data. Also, confirm key factors like age for ethical compliance.
2/Chat with a buddy who remembers you
Conversation history lets the AI remember your experiences, like a friend, and trains your personalized therapist to match your habits. Start with a new chat is also supported.
3/ Get personalized support: encouragement or solutions
The conversation provides either warm encouragement or solutions based on the AI's personality. Extremely negative emotions are detected to offer additional support.
What’s the potential users’ current pain?
What can we learn?
User research
Persona
To ensure that all team members understand our target audience and to improve our pitching, we collaboratively developed personas.
Current pains and our opportunities
Analyze the paths these potential users take when looking for a therapist and the difficulties they encounter. This will help us identify potential opportunities for our product.
Stakeholder Consensus
Align rules for integrating responsible AI
Now that we’ve defined the product’s goals and target audience, we reviewed the ethical considerations and important aspects of using responsive AI during our kickoff meeting before moving into the design phase.
The process is never linear
Ideation
Create distinct therapy personas based on user needs
Based on potential user needs, I proposed to create several appealing AI therapist personas. I wrote descriptions for each. The backend engineers used these descriptions to shape the speaking styles and chat guidance for the AI models, while I created visual representations using Midjourney.
Define
Information architecture
During the initial meetings, we established the information architecture, but it was quickly overhauled due to time constraints. Many branches were cut, leaving only the mvp scope to ensure functionality. Interestingly, some features, like the “History log”, were suggested by developers and added after group discussions.
User flow
After several iterations, our final ideal user flow is as follows.
Final version of user flow
Challenges
Implement a feedback loop at the last minutes
As the deadline neared, usability tests showed users wanted to provide feedback on the model, a typical feature in responsive AI. We planned to enhance the feedback loop in the final week, but the tight timeline and unfamiliarity with Android made implementation challenging.
Fit for user behavior
Occupies the width and reduces dialog box space
No screen space usage
Boosts user engagement
Backend development may lag behind
We ultimately chose Option A, but faced some challenges during development. In the end, we reached a compromise due to time constraints. Through effective communication and teamwork, we successfully completed this iteration together.
Reflection
01.Designing personified AI
We realized that users seeking emotional support need empathetic helpers, while those needing practical assistance require efficient solutions. This highlights the difference between the wise owl and the warm dog.
02.AI and human interaction
Implementing staged and gradual responses can reduce cognitive load for users and increase their trust in the AI.
03. Focus on MVP features
During the project sprint, we prioritized ensuring that MVP features worked, leading to the decision to drop voice conversion and profile features.
04. Collaboration with developers
Close collaboration with the development team was essential. We modified designs based on technical constraints but maintained the core UX flow as a priority.