RL Powered Mental Health Support Bot

• 

  User Input UI

• 

  Full UI

• 

  Demo Screenshot

🧠 Inspiration

Mental health care is essential but still inaccessible to many. We wanted to explore whether AI could gently guide users toward emotional wellbeing — not with generic tips, but through personalized, learned support. Reinforcement learning gave us a way to model helpful decision‑making, while GPT enabled rich emotional understanding.

💬 What it does

Our chatbot listens to the user, detects their mood using GPT‑3.5, and selects a wellness strategy using a PPO‑trained reinforcement learning agent. It also:

• Maintains conversation context across turns
  
• Tracks mood score trends visually
  
• Remembers suggestions given to avoid repetition
  
• Runs live on Hugging Face Spaces with an intuitive Gradio UI
  

🔧 How we built it

1. Created a custom Gym environment to simulate emotional state and reward improvements
   
2. Trained a PPO agent using stable‑baselines3
   
3. Used OpenAI GPT‑3.5 to classify mood from natural text input
   
4. Mapped moods to scores (e.g., sadness = 0.3, joy = 0.9)
   
5. Integrated everything into a Gradio chatbot interface
   
6. Deployed publicly via Hugging Face Spaces
   

🧱 Challenges we ran into

• Git LFS issues while pushing trained models
  
• GPT’s mood classification sensitivity required prompt tuning
  
• PPO agent needed careful reward shaping to perform well
  
• Visualizing mood charts in Gradio without crashes
  
• Handling multi‑turn conversations and maintaining session memory
  

🏆 Accomplishments that we’re proud of

• Built and deployed a live mental health agent powered by reinforcement learning
  
• Achieved real‑time mood detection + PPO suggestion integration
  
• Created a meaningful use case combining language models and RL
  
• Made the interface user‑friendly and expressive with emoji, charts, and chat history
  

📚 What we learned

• How to design a Gym environment for non‑game, human‑centered goals
  
• Best practices for deploying ML apps on Hugging Face Spaces
  
• How GPT can be used not just for generation, but semantic classification
  
• Combining symbolic reward learning with fuzzy language input
  

🚀 What’s next for RL‑Powered Mental Health Support Bot

• Add 👍 / 👎 feedback for learning from real users
  
• Use GPT‑3.5 to generate full empathetic responses
  
• Store and visualize long‑term mood history per user
  
• Add offline fallback using a local Hugging Face model
  
• Integrate external wellness APIs like Calm or Headspace
  
• Open‑source the project for research and community use
  

Built With

• gradio
• gymnasium
• huggingface
• numpy
• openai-api
• python
• stable-baselines3