Recommendation Methodology

Each recommendation request is built from inputs you provide:

• Cultural cuisine selection (e.g., West African, Caribbean, Mediterranean, South Asian, East Asian, Latin American, Middle Eastern, Southern U.S., Pan-African Diaspora).
• Goal profile: muscle gain, fat loss, or balanced maintenance.
• Household composition (Family Mode) and per-person preferences.
• Allergies, intolerances, and ingredient exclusions.
• Optional macronutrient targets (protein floor, carbohydrate ceiling, calorie band).

Every generated meal is constrained to satisfy two non-negotiable rules before it is shown to you:

• High-protein floor. Each meal targets a protein density appropriate to the user's goal profile (typically 30g+ per main meal for adults).
• Low-glycemic ceiling. Carbohydrate sources are biased toward low-to-moderate glycemic-index foods; high-GI staples are routed through the substitution graph (see Section 4).

These constraints are enforced on the server, not by the language model alone.

For each supported cuisine we maintain a curated prompt template that encodes signature ingredients, cooking methods, regional dish archetypes, and culturally meaningful flavor structures. The template is sent, together with your inputs, to a third-party large-language model via a secure gateway. The prompt library is server-only and is treated as a trade secret.

A directed graph maps culturally common but high-glycemic staples (for example, white rice, fufu, tortillas, injera) to culturally appropriate lower-glycemic equivalents, each annotated with:

• Glycemic delta — change in glycemic load vs. the original.
• Protein delta — change in protein density vs. the original.
• Cultural-fit score — how appropriate the substitution is for the chosen cuisine.

If a generated meal violates the high-protein or low-glycemic guardrails, the engine automatically applies the best-scoring culturally appropriate substitution and records the swap so you can see exactly what was changed and why.

Each finalized meal is scored 0–100 for cultural authenticity using a deterministic ingredient co-occurrence model derived from regional culinary sources. Ingredients are classified as signature, supporting, or foreign for the chosen cuisine and weighted accordingly. The score is surfaced to support transparency; it is not a judgment of any individual's heritage or cooking.

For each request the server executes, in order:

1. Validate and normalize inputs.
2. Select the cuisine-specific prompt template.
3. Call the LLM gateway to generate candidate meals.
4. Run the substitution graph to repair any guardrail violations.
5. Compute the Cultural Authenticity Index for each repaired meal.
6. Return the meals, applied substitutions, and authenticity scores to the client.

• We do not diagnose, treat, cure, or prevent any disease. See our Disclaimer.
• We do not claim a meal is medically appropriate for any specific condition.
• We do not use your personal inputs to train third-party general-purpose AI models where opt-out is available. See our Privacy Policy.
• We do not store or process Protected Health Information (PHI) under HIPAA today.

Generated meals are suggestions based on the inputs you provide. Nutritional values are estimates. Real-world outcomes depend on portion sizes, ingredient quality, preparation, adherence, and individual physiology. Always consult a qualified healthcare professional before making significant dietary changes.

The methodology, prompt library, substitution graph, and authenticity model are updated over time. Material changes will be reflected on this page.