AI-Powered Personalized Health Assistant

Project Overview

The AI-Powered Personalized Health Assistant leverages advanced Generative AI techniques to provide users with personalized fitness and nutrition recommendations. Utilizing fine-tuned large language models and structured workflow management, the assistant offers individualized health guidance tailored to diverse user profiles and goals.

---

Team

• Kanika Rawat
• Vasu Vinodbhai Bhut
• Krupali Gunvantbhai Tejani

---

Technologies Used

• OpenAI API: Access and fine-tuning of language models
• Pandas, NumPy: Data manipulation
• Scikit-learn: Machine learning utilities
• LangChain, LangGraph: Workflow management
• Streamlit: User interface development

---

Features

Personalized Recommendations

• Tailored workout routines
• Customized dietary guidelines
• Continuous engagement and motivation through interactive messaging

Structured Workflow

• Node Functions:
  • generate_workout_plan
  • generate_diet_plan
  • format_response
• Directed Graph-based workflow ensuring coherent response integration
• State management via TypedDict for workflow consistency

Prompt Engineering & Fine-Tuning

• Techniques Explored:
  • Zero-Shot Prompting
  • Few-Shot Prompting
  • Chain-of-Thought (CoT) Prompting
• Fine-tuning executed on GPT-3.5-Turbo with custom dataset, enhancing health domain relevance

Robust Security Implementation

• PromptDefender Class:
  • Regular expression-based pattern matching
  • Context-sensitive validation
  • Prevention against encoding-based attacks
• Multi-Stage Security Pipeline:
  • Input sanitization
  • Secure query processing
  • Output validation

User Interface with Streamlit

• User Profile Management
  • Physical attributes, dietary preferences, health goals
• Interactive Chat Interface
  • Conversation tracking and response rating
• Admin Dashboard
  • Metrics monitoring including response times, user ratings, and query volumes

---

Security and Ethical Considerations

• Anonymous interactions (no personal data stored)
• Stateless and session-based responses
• Explicit disclaimer against providing medical diagnoses
• No medication or clinical treatment advice
• Ethical dataset usage (public and anonymized data)

---

Challenges and Solutions

Challenge: Fine-Tuning Data Quality

• Problem: Initial fine-tuning was inconsistent due to limited quality and diversity of data.
• Solution: Enhanced data generation with broader demographics, expert-reviewed responses, and rigorous data validation.

Challenge: Prompt Injection Vulnerabilities

• Problem: System initially susceptible to prompt injection attacks.
• Solution: Developed the comprehensive PromptDefender class incorporating context-aware validation and thorough input/output sanitization.

Challenge: Response Coherence

• Problem: Disjointed workout and nutrition advice.
• Solution: Implemented structured LangGraph workflow to integrate domain-specific advice effectively.

Challenge: Performance Optimization

• Problem: Slow response times with complex queries.
• Solution: Introduced caching, optimized prompts, asynchronous processing, and fallback models.

---

Performance Metrics

• Response Quality: Enhanced significantly post fine-tuning
• Security Efficacy: Successfully blocked 91% of prompt injection attempts
• Average Response Time: Reduced to 17.90 seconds for complex queries
• User Satisfaction: Positive feedback for personalized recommendations

---

Future Vision

• Integration with wearable devices for real-time health data
• Expansion into mental health and chronic disease management
• Clinical validation to enhance reliability and impact

---

References

• Artificial Intelligence in public health nutrition, personalized AI nutritionist
• Google's Personal Health Large Language Model (PH-LLM)
• ACE Fitness insights on AI in health and fitness
• MDPI review on AI and personalized nutrition
• Evaluation of AI virtual health assistants

---

Project Repository

For complete source code and further details, visit the project repository (https://github.com/vaisu-bhut/Health-Assistant.git).

---

Project Developed for:
Prompt Engineering for Generative AI
Professor: Shirali Patel