Zdravko AI — Science-Based Nutrition

A conversational nutrition assistant that tracks four daily pillars while applying explicit safeguards for unsafe restriction, medical boundaries, supplements, and vulnerable users.

Comment

Inspiration

People increasingly use general-purpose AI for everyday nutrition questions: estimating meals, setting calorie targets, asking about protein, hydration, supplements, or weight loss. That can be useful, but nutrition advice can quickly cross into safety-sensitive territory, including unsafe restriction, disordered-eating patterns, pregnancy, chronic disease, medications, allergies, or supplement use.

Zdravko AI was built around a simple question: how can a conversational nutrition assistant stay useful for ordinary adults while knowing when not to simply comply?

What it does

Zdravko AI is a conversational assistant for practical nutrition self-monitoring. Instead of asking users to track every nutrient, it focuses on four daily pillars:

  • Calories — the most direct check on total energy intake
  • Protein — supports satiety, muscle maintenance, and a more substantial eating pattern
  • Fruit and vegetables — a practical signal for fibre, micronutrients, and dietary variety
  • Water — a simple hydration reminder

The assistant works with imperfect everyday input, such as “a bowl of cereal with milk,” and gives approximate estimates rather than pretending to be perfectly precise.

The distinctive part is the safety layer. Zdravko AI includes explicit safeguards for situations such as very low calorie targets, rapid weight loss, restrict–compensate patterns, possible disordered eating, minors, pregnancy, chronic disease, medication-related nutrition issues, allergies, and supplement requests. Depending on the case, it can answer normally, qualify and redirect to a safer target, or defer to a physician, dietitian, pharmacist, therapist, or urgent care.

How we built it

Zdravko AI was developed as an instruction-level safety layer on top of a general-purpose conversational AI system. The current version is implemented as a custom GPT with core behavioural instructions and supporting knowledge documents.

The development process followed seven stages:

  1. Literature review
  2. Source verification
  3. Prompt development
  4. Safeguard design
  5. Implementation
  6. Testing
  7. Refinement

The nutrition anchors were based on manually checked scientific and institutional sources, including energy-estimation equations, protein-intake guidance, hydration reference values, and fruit-and-vegetable recommendations. The safety layer was then tested with ordinary, ambiguous, edge-case, and high-risk prompts.

A 50-case benchmark was also created to compare default ChatGPT with the Zdravko version. The benchmark was designed to test both under-caution and over-caution: unsafe responses are failures, but unnecessary refusal or excessive referral in low-risk cases is also a failure.

Challenges we ran into

The hardest challenge was calibration. A safety-focused assistant should not simply refuse everything. It needs to remain useful for ordinary nutrition tracking while refusing or redirecting genuinely unsafe requests.

Another challenge was conversational context. A user may begin with a harmless meal-logging question and later ask for an unsafe calorie target or compensatory fasting. The assistant needs to remember relevant context without overreacting.

Supplement advice was also difficult. A categorical rule against supplement doses improves harm avoidance, but can become too cautious for routine preventive questions such as vitamin B12 in vegan diets or vitamin D in winter. This became one of the clearest design lessons.

Accomplishments that we're proud of

Zdravko AI is more than a generic nutrition chatbot. It combines a simple four-pillar tracking model with explicit safety boundaries and a benchmark for testing whether those boundaries change behaviour.

I am especially proud that the project does not treat safety as only refusal. The goal is safer usefulness: answer when appropriate, qualify when needed, and defer when the request goes beyond the assistant’s safe scope.

What we learned

Writing a safeguard into a prompt is not enough. It has to be tested as behaviour across realistic conversations, including user pressure, incomplete information, and subtle risk signals.

The project also showed that excessive caution is a real failure mode. A nutrition assistant that refuses ordinary low-risk questions may be safer in a narrow sense, but less useful in practice. Good health AI needs both harm avoidance and calibration.

What's next

The next step is to turn Zdravko AI from a custom-GPT prototype into a more reproducible product with a clearer technical interface, versioned safeguards, and systematic evaluation.

Planned improvements include:

  • a larger benchmark with independent multi-turn scenarios;
  • clinician, dietitian, pharmacist, and eating-disorder-specialist review;
  • more granular risk tiers for supplement questions;
  • a lightweight web interface;
  • clearer user controls for relaxed, moderate, or precise tracking.

Built With

  • ai-safety
  • benchmarking
  • chatgpt
  • conversational-ai
  • custom-gpt
  • digital-health
  • evaluation
  • generative-ai
  • gpt
  • healthcare
  • llm
  • nutrition
  • prompt-engineering
  • responsible-ai
Share this project:

Updates

Submission history