Zdravko AI

• 

  Zdravko_AI_Showcase_Conversation_page-0002

• 

  Zdravko_AI_Showcase_Conversation_page-0001

• 

  Zdravko_AI_Showcase_Conversation_page-0003

• 

  Zdravko_AI_Showcase_Conversation_page-0004

• 

  Zdravko_AI_Showcase_Conversation_page-0006

• 

  Zdravko_AI_Showcase_Conversation_page-0005

• 

  Zdravko_AI_Configuration

• 

  Zdravko_AI_logo

• 

  Figure1_safety_score_distribution

• 

  Zdravko_AI_Architecture.png

Inspiration

Most nutrition apps require detailed food logging. General-purpose AI makes tracking easier, but health-related conversations can involve unsafe restriction, eating disorders, medication interactions, and other risks.

Zdravko AI explores whether a general-purpose chatbot can provide practical, evidence-informed nutrition guidance while maintaining appropriate safety boundaries.

What it does

Zdravko adds two instruction-based layers to an existing conversational AI:

1. Nutrition layer

Tracks four daily pillars:

• Calories — energy balance
• Protein — muscle maintenance and satiety
• Fruit & vegetables — dietary variety and micronutrients
• Water — hydration

Users describe meals naturally, and Zdravko estimates their intake without requiring precise measurements.

2. Safety layer

Recognizes potentially unsafe requests involving restrictive eating, vulnerable users, medical conditions, medications, allergies, and supplements.

Depending on risk, it provides guidance, introduces appropriate precautions, or recommends professional care.

How we built it

Zdravko is a custom GPT, not a newly trained model. It combines structured instructions, reference documents, and evidence-informed nutritional recommendations.

Development involved literature review, source verification, prompt engineering, safeguard design, testing, and refinement.

Challenges and results

The central challenge was balancing harm prevention with usefulness. Excessive caution can make an assistant unnecessarily restrictive.

In an exploratory 50-case evaluation, Zdravko received higher automated safety scores in 13 cases, tied in 34, and scored lower in three.

The study identified context-related failure patterns and overcautious supplement responses. However, model differences, dependent observations, and a single automated evaluator prevent causal or clinical safety claims.

What's next

We plan to improve risk-calibrated safeguards, conduct independent multi-turn evaluations with clinical experts, and develop a more reproducible interface.

Zdravko AI is a research prototype, not a medical device or substitute for professional healthcare.

Built With

• ai-safety
• benchmarking
• chatgpt
• conversational-ai
• custom-gpt
• digital-health
• generative-ai
• healthcare
• llm
• nutrition
• prompt-engineering
• responsible-ai