Embrace AI

Embrace AI revolutionizes stress management by leveraging TiDB’s fast serverless tech and vector search for real-time insights, while simplifying AI adoption for engineers.

Comment

Inspiration

In today’s fast-paced world, many wellness apps fail to provide personalized emotional support. Embrace AI addresses this gap by using TiDB’s serverless infrastructure and advanced vector search, delivering real-time, tailored stress management solutions.

What It Does

Embrace AI is a personalized stress management platform powered by an AI swarm that processes facial images through a custom CNN and integrates location data using text embeddings from Huggingface, generating vectors stored in TiDB for accurate context-aware recommendations. Our RL model then recommends mindfulness exercises based on the user’s emotion and context, refining over time using Q-learning with hybrid on-policy and off-policy methods, leveraging TiDB’s vector search for speed.

How We Built It

Our tech stack integrates:

  • Custom CNN for facial analysis (trained on 15,000 images).
  • Text Embedding from Huggingface for vectorizing text data.
  • Q-Learning RL Model for personalized exercise recommendations:
    • On-Policy: TiDB’s Cosine Similarity function for vector matching.
    • Off-Policy: Q-value recommendations.
  • Python for core development (TensorFlow, Gymnasium, NumPy, Pandas).
  • Docker Containers for deployment.
  • TiDB Serverless for fast, scalable storage and retrieval.
  • Google Cloud Run for serverless operations.
  • Dash Framework with Plotly for interactive visualizations.
  • Jupyter Notebooks and VS Code for development.

Teamwork played a key role in bringing this all together into a cohesive and scalable solution.

Challenges We Ran Into

  • First time building an AI swarm and RL model, overcoming steep learning curves.
  • Debugging gradient issues in model training.
  • Learning vector search and ensuring correct formatting.
  • Acquiring new skills like Docker and Google Vertex AI.
  • Managing multiple APIs and Docker containers.
  • Integrating complex models while maintaining usability.
  • Limited data, necessitating creative use of generative AI with scripting.
  • Working through outdated documentation for essential tools.

What We Learned

  • Seamlessly integrating diverse AI technologies into a single application.
  • Handling real-time data with serverless databases like TiDB.
  • Prioritizing user-centered design for both functionality and education.
  • Gaining in-depth knowledge of vector search, RL, and AI collaboration.
  • Mastering serverless tech (TiDB, Google Cloud Run) and Docker.

What's Next for Embrace AI

We plan to:

  • Expand emotional recognition to include voice tone analysis.
  • Add more emotion categories and allow custom image/location inputs.
  • Integrate with wearables (e.g., Apple Watch) for quicker engagement.
  • Incorporate contextual factors like weather and time for better personalization.
  • Enhance feedback loops with deep RL for more precise recommendations.

Built With

Share this project:

Updates