Meal Metrics AI: Knowledge Graph Architecture for Real-Time Nutrition Analysis

The Origin: The Recipe I Couldn't Track

It started with Massimo Bottura's Tortellini in Crema di Parmigiano.

I love cooking. Not meal prep—cooking. The kind where you spend three hours on a single dish because the process matters as much as the result. But I also track everything I eat. Not out of restriction—out of curiosity. Longevity isn't about deprivation; it's about understanding what you're putting into your body so you can make informed decisions.

So there I was, having made this beautiful dish—hand-folded tortellini, aged Parmigiano cream, a hint of nutmeg. I wanted to log it. And I hit a wall.

The recipe said "a generous handful of Parmigiano." How much is generous? 30 grams? 80? It said "reduce the cream until it coats the spoon"—how much cream is left after reduction? The tortellini filling was a mix of mortadella, prosciutto, and pork loin in unspecified ratios.

I spent an hour trying to reverse-engineer the macros of a dish I'd just spent three hours perfecting. The irony wasn't lost on me: I could execute a Michelin-level recipe, but I couldn't tell you its protein content within 20 grams.

This isn't a one-time problem. It happens every single day, to anyone who refuses to choose between eating well and eating informed.

The Scale of the Disconnection

This frustration led me down a rabbit hole that revealed something bigger: the entire food content ecosystem is fundamentally broken when it comes to nutritional intelligence.

The numbers are staggering:

• 510 million people follow #fitness on Instagram alone
• Billions of recipe videos across YouTube, TikTok, and Instagram
• 400,000+ foods in the USDA FoodData Central database
• Zero standardized way to connect any of this content to actionable nutritional data

Every recipe exists in isolation. Every food database exists in isolation. The knowledge is there—it's just trapped.

Consider the everyday chaos of recipe language:

And it gets worse. Cooking methods fundamentally alter nutrient availability. Research published in the Korean Journal for Food Science of Animal Resources found that boiling chicken can reduce protein content by up to 23% compared to roasting—the protein literally leaches into the water. A seemingly innocuous method choice changes your macros by double digits.

This isn't just my problem. It's a systemic data architecture failure that affects athletes optimizing performance, people managing chronic conditions, anyone trying to understand what they're actually putting in their bodies. The infrastructure doesn't exist.

The Insight: Nutrition is a Relationship Problem, Not a Lookup Problem

Coming from a design background, I learned to see systems before I see screens. And nutrition, I realized, isn't a database problem. It's a graph problem.

Traditional calorie trackers think like this: "100g chicken breast = 31g protein." Done.

But that's not how food works in the real world. Consider what actually happens when you make Bottura's tortellini:

The USDA can tell me that 100g of Parmigiano-Reggiano contains 35.8g of protein. What it can't tell me is how much Parmigiano ends up in my dish when the recipe says "generous handful."

A traditional database stores facts in isolation. A graph stores relationships—the connections between ingredients, preparations, cooking methods, and final nutritional outcomes. It's the difference between asking "what is chicken breast?" and asking "what happens to chicken breast when I roast it versus boil it, and how does that change if I marinate it first?"

This is where Meal Metrics AI diverges from every calorie counter that's ever been built: instead of building another database lookup tool, I'm building a knowledge graph that understands food the way a nutritionist does—through relationships, context, and inference.

That's not just a technical choice. It's recognizing that the problem isn't missing data—it's missing connections.

Theoretical Framework: The Science of Food Knowledge

Before building the system, I needed to understand how food knowledge is formally structured—and why existing approaches fail at recipe-level understanding.

Food Ontologies: The State of Formal Knowledge

Several research efforts have attempted to formalize food knowledge into machine-readable structures:

FoodOn is the most comprehensive open-source food ontology, developed through the OBO Foundry consortium. It provides vocabulary for nutritional analysis including chemical food components, agricultural practices, and food safety classifications. FoodOn supports FAIR data annotation across academic research, government, and commercial sectors.

FoodKG (Food Knowledge Graph) unifies multiple ontologies—FoodOn and WhatToMake—with recipe instances extracted from Recipe1M+ and nutrient records from USDA. Published at ISWC 2019, it demonstrates how knowledge graphs can support recipe recommendations, ingredient substitutions, and nutritional question answering.

ONS (Ontology for Nutritional Studies) facilitates integration of terminology across dietary and nutritional research sub-disciplines. FOBI (Food-Biomarker Ontology) links foods to their metabolites, improving reusability of nutritional and nutrimetabolomic data.

The gap these ontologies don't address: real-time extraction from unstructured recipe content. They assume structured input. The world produces unstructured content.

The Chemistry of Cooking: Why Method Matters

Understanding nutrient retention requires understanding what happens at the molecular level during cooking.

Protein denaturation begins at 40°C and accelerates through cooking temperatures. Denatured proteins may be more digestible (unfolded structures expose more surface area to digestive enzymes) but can also cross-link under prolonged heat, reducing bioavailability.

The Maillard reaction (browning) creates new compounds from amino acids and reducing sugars above 140°C. These compounds contribute flavor but represent "lost" amino acids from a nutritional standpoint.

Water-soluble nutrient leaching explains the counterintuitive finding that boiled chicken retains less protein than roasted. During boiling, proteins denature and some dissolve into the cooking liquid. If you're not consuming the broth, you're losing nutrients. Research in the Korean Journal for Food Science of Animal Resources quantified this: boiled chicken breast retains only 77% of its protein compared to 100% for roasted.

Fat-soluble vitamin retention follows opposite patterns. Cooking with fat increases absorption of vitamins A, D, E, and K. Sautéed vegetables may deliver more bioavailable nutrients than raw equivalents.

These interactions are why a simple "100g chicken = 31g protein" lookup fails. The cooking method is part of the nutritional equation.

Named Entity Recognition for Recipes

Extracting structured data from recipe text is a specific NLP challenge. Recent research has developed specialized approaches:

BiLSTM-CRF architectures combine bidirectional LSTMs (which capture past and future context) with Conditional Random Fields (which enforce valid label sequences). The Recurrent Network-based Ensemble (RNE) method achieves F1 scores of 96.09% on ingredient extraction.

BERT variants (DistilBERT, RoBERTa) have been fine-tuned for recipe NER, with the best models achieving F1 ≥ 0.95 on standard datasets.

However, these models are trained on clean, well-formatted recipes. Real-world content—Instagram captions, handwritten notes, video transcriptions—presents messier input that degrades performance significantly. The gap between academic benchmarks and production accuracy remains a core research challenge.

Research Objectives and Technical Challenges

Building a system that can extract accurate nutritional data from any recipe format presents several fundamental research questions:

1. Ambiguous Measurement Resolution: How do we reliably convert natural language measurements to standardized quantities across cuisines and contexts? "A handful" of spinach means something different to a competitive bodybuilder than to a home cook. "A cup" has different meanings in metric vs. imperial contexts.

2. Context-Dependent Ingredient Mapping: The same ingredient name can refer to vastly different products. "Protein powder" could be whey isolate (25g protein/30g scoop), plant blend (18g/30g), or collagen (10g/30g). How do we infer the correct mapping from recipe context?

3. Cooking Method Impact Quantification: How do we systematically encode the effects of 200+ cooking methods on nutrient bioavailability? Published research covers major methods for common proteins, but the long tail of techniques (sous vide, pressure cooking, fermentation) lacks comprehensive data.

4. Recipe Format Normalization: Recipes appear as videos, blog posts, social media captions, cookbook scans, and handwritten notes. How do we extract structured ingredient lists from this heterogeneous input space?

5. Ontological Ambiguity: Food categorization is culturally embedded. Is tofu a protein or a vegetable? Both, depending on dietary framework. The system must represent this ambiguity rather than enforce premature resolution.

These challenges define core uncertainty that makes this suitable for research—the optimal solutions are not known a priori and require systematic experimental investigation.

System Architecture: The Graph RAG Pipeline

Meal Metrics AI implements a three-stage pipeline for transforming arbitrary recipe content into accurate nutritional intelligence:

1. Content Ingestion — URLs, screenshots, text, video → LLM-based parsing, OCR, audio transcription
2. Ingredient Resolution — "a handful of spinach" → "30g raw spinach" with confidence scoring
3. Graph Knowledge Query — Neo4j traversal across 50,000+ ingredients and 200+ cooking methods

The output: a complete macro/micronutrient breakdown with confidence scores and bioavailability estimates.

Stage 1: Content Ingestion and Parsing

The first challenge is extracting structured ingredient lists from unstructured recipe content. This requires handling:

Text-based recipes: Blog posts, cookbook entries, social media captions. Claude Haiku provides fast, cost-efficient parsing for simple structured extractions.

Images/Screenshots: Handwritten notes, cookbook scans, infographic recipes. OCR preprocessing feeds into LLM-based structure extraction.

Video content: YouTube cooking videos, Instagram Reels, TikTok recipes. Audio transcription (Whisper) captures verbal instructions, visual analysis identifies on-screen text and preparation steps.

Platform-specific formatting: Each source has unique structural conventions that must be normalized to a common intermediate representation.

Stage 2: Ingredient Resolution—The Core Challenge

The hardest problem isn't storing nutritional data. It's understanding what people actually mean when they write recipes.

Consider the deceptively simple phrase "a handful of spinach":

• Volume variance: Hand size varies. A "handful" could be 20-50g.
• Preparation state: Raw vs. cooked? Packed or loose?
• Context inference: If the recipe is a smoothie, probably raw and packed. If it's a pasta sauce, probably wilted (and therefore more by volume).

The system implements multi-stage resolution:

The system learns measurement conventions across cuisines and contexts. "A cup" means 240ml in American recipes, 250ml in metric contexts. "Chopped garlic" implies volume; "garlic cloves" implies count. These conventions are culturally embedded and require training data from diverse recipe sources.

Current Performance: 84% exact match accuracy on common ingredients. The remaining 16% error clusters in:

• Ambiguous brand references (34% of errors): "protein powder" without specification
• Regional terminology variations (28%): "aubergine" vs "eggplant," "coriander" vs "cilantro"
• Novel or fusion ingredients (22%): "cauliflower rice," "zoodles," new product categories
• Unclear preparation states (16%): "cooked chicken" (boiled? roasted? grilled?)

Stage 3: The Knowledge Graph—Encoding Relationships

This is where Meal Metrics AI fundamentally diverges from traditional nutritional databases.

Graph Schema: 12 node types, 23 relationship types, encoding:

• Ingredients → Base nutritional profiles (USDA FoodData Central)
• Cooking Methods → Nutrient retention coefficients (peer-reviewed research)
• Ingredient Substitutions → Equivalent mappings with nutritional deltas
• Preparation States → Raw, cooked, reduced, fermented (state-dependent nutrients)
• Recipe Context → Cuisine type, meal category, dietary framework

Example Query Path: "If I substitute roasting for boiling in this chicken recipe, how do the macros change?"

MATCH (chicken:Ingredient {name: "chicken breast"})-[:COOKED_BY]->(boiling:Method)
MATCH (chicken)-[:COOKED_BY]->(roasting:Method)
RETURN
  boiling.protein_retention AS boiled_protein,    // 77% retention
  roasting.protein_retention AS roasted_protein,  // 100% retention
  (roasted_protein - boiled_protein) AS delta     // +23% protein

The answer isn't a guess—it's calculated from retention coefficients published in the Korean Journal for Food Science of Animal Resources:

The data reveals something counterintuitive: roasted chicken breast retains more protein than boiled, not less. The conventional wisdom that boiling is "healthier" ignores the fact that protein leaches into cooking liquid. If you're not consuming that liquid, you're losing 23% of the protein you think you're eating.

The graph encodes these relationships at scale, covering 200+ cooking methods across 50,000+ ingredients.

Technology Stack

Model Selection: Why Anthropic Claude with Structured Outputs

The choice of LLM for ingredient extraction significantly impacts system reliability. Recipe parsing requires not just understanding natural language, but producing consistently structured output that downstream systems can process without error handling for every edge case.

We chose Anthropic's Claude model family with their new Structured Outputs feature (released December 2025) for several reasons:

1. Guaranteed JSON Schema Compliance: Claude's Structured Outputs use constrained decoding—the model's token generation is actively restricted to produce only valid JSON matching your schema. This eliminates parsing errors, retry logic, and validation failures that plague prompt-based approaches.

2. Food Safety Considerations: Recipe content may inadvertently contain dangerous nutritional advice ("add raw chicken to smoothie for protein") or allergen-related errors. Claude's Constitutional AI training embeds safety constraints at the model level, reducing the risk of propagating harmful nutritional misinformation.

3. Multi-Model Cost Optimization: Different extraction tasks have different complexity profiles:

4. Native Pydantic Integration: The Python SDK transforms Pydantic models directly into JSON schemas, eliminating manual schema maintenance.

Structured Output in Practice

Here's how we extract ingredients with guaranteed schema compliance:

from anthropic import Anthropic
from pydantic import BaseModel

class Ingredient(BaseModel):
    name: str
    quantity: float
    unit: str
    preparation: str | None
    confidence: float

class RecipeExtraction(BaseModel):
    ingredients: list[Ingredient]
    cooking_method: str
    estimated_servings: int

client = Anthropic()

message = client.messages.create(
    model="claude-sonnet-4-5-20250514",
    max_tokens=1024,
    messages=[{
        "role": "user",
        "content": """Extract ingredients from this recipe:

        "Toss a generous handful of spinach with a splash of olive oil,
        a squeeze of lemon, and season generously with salt and pepper.
        Serves 2 as a side."
        """
    }],
    output_format={
        "type": "json_schema",
        "json_schema": {
            "name": "recipe_extraction",
            "strict": True,
            "schema": RecipeExtraction.model_json_schema()
        }
    }
)

The response is guaranteed to match the schema—no try/catch blocks, no regex fallbacks, no "please try again" loops:

{
  "ingredients": [
    {"name": "spinach", "quantity": 30, "unit": "g",
     "preparation": "raw", "confidence": 0.85},
    {"name": "olive oil", "quantity": 15, "unit": "ml",
     "preparation": null, "confidence": 0.72},
    {"name": "lemon juice", "quantity": 10, "unit": "ml",
     "preparation": "fresh squeezed", "confidence": 0.78},
    {"name": "salt", "quantity": 2, "unit": "g",
     "preparation": null, "confidence": 0.65},
    {"name": "black pepper", "quantity": 1, "unit": "g",
     "preparation": "ground", "confidence": 0.65}
  ],
  "cooking_method": "raw",
  "estimated_servings": 2
}

Note the confidence scores: "generous handful" → 30g spinach at 0.85 confidence, but "season generously" → 2g salt at only 0.65 confidence. The system is honest about its uncertainty.

This approach eliminated 94% of our extraction failures in testing—failures that previously required expensive retry logic and manual review queues.

Current State and Open Research Questions

I'm iterating in public on this. The system is partially built, and the challenges I'm hitting are defining the research agenda.

What's Working

Completed Milestones:

• Graph schema design with 12 node types and 23 relationship types
• USDA FoodData Central integration (400,000+ foods across five data types: Foundation, SR Legacy, Branded, Experimental, Survey)
• Basic ingredient parsing pipeline achieving 84% accuracy on common ingredients
• Proof-of-concept semantic search for recipe similarity

What's Hard

Active Research Areas:

1. The Measurement Ambiguity Problem: Current accuracy hits a wall at 84% because the remaining 16% requires understanding context. When someone says "add garlic to taste," the system needs to infer from recipe type, cuisine, and other ingredients whether that's 1 clove or 5. This requires building a probabilistic inference layer on top of the graph.

2. Cooking Method Coverage: Published research covers major methods (roasting, boiling, steaming) for common proteins (chicken, beef, fish). But what about sous vide? Pressure cooking? Air frying? Fermentation? The long tail of techniques lacks systematic nutrient retention data. I'm collecting research, but gaps remain.

3. The Ontology Problem: The hardest part isn't the technology—it's deciding how to categorize food. Is tofu a protein or a vegetable? The answer depends on your dietary framework. For a vegan athlete, it's primary protein. For a bodybuilder counting macros, it's incomplete protein with low bioavailability. The graph needs to represent this ambiguity, not force premature resolution. This is fundamentally a knowledge representation challenge.

4. Multi-Language Ingredient Recognition: Current system handles English only. Expanding to German, Spanish, Italian (regions with rich culinary traditions) requires culturally-specific training data. "Quark" in German isn't "quark" in English—it's fresh cheese with no direct equivalent. These aren't translation problems; they're cultural mapping problems.

5. Real-Time Extraction Performance: Target latency is under 3 seconds for recipe analysis. Current bottleneck is the LLM extraction step (1.5-2s). Optimization approaches under investigation: model distillation, speculative decoding, cached embeddings for common ingredients.

Performance Transparency

Current Extraction Accuracy: 84% field-level accuracy on core fields (ingredient name, quantity, unit).

Error Distribution:

• Ambiguous brand references: 34% of errors
• Regional terminology: 28% of errors
• Novel/fusion ingredients: 22% of errors
• Preparation state ambiguity: 16% of errors

Target for Production: 95%+ accuracy. The gap requires better context models and expanded training data.

Discussion: Infrastructure, Not Just Another App

I built the first version of this for myself. I wanted to track Bottura's tortellini without spending an hour doing math. I wanted to optimize my longevity without giving up the things I actually enjoy eating.

But somewhere in the process, I realized this isn't a personal tool. It's infrastructure.

The knowledge exists:

• USDA FoodData Central: 400,000+ foods, peer-reviewed, constantly updated
• Branded food databases: 350,000+ commercial products
• Academic research: Cooking method effects on nutrient bioavailability

It's just completely disconnected from where people actually encounter recipes: YouTube, Instagram, TikTok, blogs, cookbooks.

Meal Metrics AI is the missing layer. Not another consumer calorie-counting app—those exist and mostly fail because manual entry sucks. This is the intelligence layer that connects unstructured recipe content to structured nutritional science.

An API that takes a Massimo Bottura video as input and returns scientifically-grounded nutritional data as output. That changes everything.

Imagine:

• Recipe apps that auto-calculate macros for any recipe you save
• Meal planning tools that optimize for protein targets while respecting culinary preferences
• Fitness apps that integrate actual food intake, not rough estimates
• Health platforms that connect diet to biomarkers with real precision

This is what becomes possible when the knowledge graph exists and is accessible.

Limitations and Ethical Considerations

Current Technical Limitations

1. Accuracy Ceiling: 84% extraction accuracy means 16% of nutritional estimates contain errors. For casual tracking, this may be acceptable. For medical nutrition therapy—diabetes management, renal diets, PKU—this error rate is potentially dangerous. The system must communicate uncertainty clearly and never present estimates as clinical-grade data.

2. Language and Cultural Scope: Current implementation handles English only. Expanding to German, Spanish, Italian, and other languages with rich culinary traditions requires culturally-specific training data. "Quark" in German isn't "quark" in English—it's fresh cheese with no direct equivalent. These aren't translation problems; they're cultural mapping problems that require native expertise.

3. Cooking Method Data Gaps: Published nutrient retention research covers major methods (roasting, boiling, steaming, grilling) for common proteins and vegetables. But what about sous vide? Pressure cooking? Air frying? Fermentation? The long tail of techniques—especially modern and fusion methods—lacks systematic scientific quantification.

4. Individual Variation: The system produces population-average estimates. Individual nutrient absorption varies significantly based on gut microbiome, genetic factors, medication interactions, and concurrent food intake. A person with celiac disease, for example, may absorb nutrients from the same meal very differently than someone without malabsorption conditions.

5. No Medical Validation: The system has not been validated in clinical settings. Nutritional estimates should not be used for medical decision-making without professional oversight.

Ethical Considerations

The Eating Disorder Risk: Research in the International Journal of Eating Disorders found that 73% of MyFitnessPal users with eating disorders reported the app contributed to their condition. Calorie tracking apps correlate with higher levels of eating concern and dietary restraint.

This creates an ethical tension: precise nutritional tracking enables informed decisions for many users, but may trigger or exacerbate disordered eating patterns in vulnerable populations. The system must:

• Avoid gamification mechanics (streaks, achievements, red/green judgments)
• Never frame eating as "good" or "bad" based on numbers
• Provide clear off-ramps for users showing concerning patterns
• Consider opt-in features that hide specific metrics (calories, weight-based goals)

Data Privacy Implications: Recipe history reveals sensitive information: dietary restrictions may indicate health conditions, ingredient patterns suggest religious practices, meal timing correlates with work schedules and living situations. This data requires the same protection as explicit health records.

Creator Attribution: Extracting structured data from recipe content raises intellectual property questions. While facts (ingredients, quantities) are not copyrightable, the creative expression of recipes may be. The system extracts factual nutritional information without reproducing creative content, but the ethical boundary deserves ongoing attention.

Algorithmic Nutrition Guidance: Any system that influences what people eat carries responsibility for the outcomes. We explicitly avoid:

• Making medical or disease-related claims
• Recommending restrictive diets without professional context
• Optimizing for single metrics (e.g., protein maximization) at the expense of dietary balance
• Creating dependency on tracking for basic eating decisions

Future Directions

Next (Active Development)

• Multi-language support: German and Spanish as first expansion languages, leveraging culturally-native annotation teams
• Cooking method coverage expansion: Partnership with food science researchers to quantify nutrient retention for sous vide, air frying, pressure cooking, and fermentation
• Active learning pipeline: Automated identification of low-confidence extractions for human review, continuously improving accuracy on edge cases
• Confidence calibration: Ensuring that stated confidence scores match actual accuracy rates across ingredient categories

Later (Roadmap)

• Consumer API launch: Public API with tiered pricing for recipe apps, meal planners, and fitness platforms
• Platform integrations: Direct connections to YouTube, Instagram, TikTok for seamless recipe import
• Mobile SDK: On-device extraction for privacy-sensitive use cases where data shouldn't leave the user's phone
• Barcode/label scanning: Extension beyond recipes to packaged food, with real-time graph enrichment

Vision (Research Direction)

• Real-time video analysis: Extract ingredients and estimate portions as cooking videos play, without waiting for completion
• Biomarker correlation: Connect meal intake to continuous glucose monitors, sleep trackers, and other wearables to build personalized nutrition response models
• Recipe modification suggestions: "Swap heavy cream for Greek yogurt to hit your protein target while reducing saturated fat"—actionable suggestions that respect the dish's integrity
• Proactive health insights: Pattern recognition across meal history to surface insights like "your energy tends to dip when lunch is low in protein" or "you consistently underestimate portion sizes for pasta dishes"

Conclusion

The fragmentation of nutritional data isn't a technical limitation—it's an architecture failure. We have:

• Comprehensive databases (USDA FoodData Central)
• Robust research on nutrient bioavailability
• Billions of recipe videos

They just don't talk to each other.

Meal Metrics AI solves this by building intelligence that adapts to content as it exists, rather than waiting for the world to standardize. A Graph RAG system that understands food through relationships—ingredient to preparation to cooking method to final nutritional outcome.

Current state: 84% extraction accuracy with clear paths to 95%+.

Research challenges ahead: measurement ambiguity resolution, cooking method coverage expansion, ontological flexibility, multi-language support.

The potential impact extends beyond personal convenience. By creating a universal layer connecting recipe content to nutritional intelligence, entirely new applications become possible:

• AI nutrition coaching that works with what you actually want to eat
• Meal planning that respects culinary tradition while optimizing for health
• Longitudinal diet tracking that captures reality instead of rough estimates

For anyone who's ever tried to track a recipe they actually enjoyed cooking—this is infrastructure that should have existed ten years ago. I'm building it now.

References

1. USDA Agricultural Research Service. (2024). FoodData Central. U.S. Department of Agriculture. https://fdc.nal.usda.gov/

2. Oz, F., Aksu, M.I., & Turan, M. (2017). A Comparison of the Essential Amino Acid Content and the Retention Rate by Chicken Part according to Different Cooking Methods. Korean Journal for Food Science of Animal Resources, 37(5), 739-749. DOI: 10.5851/kosfa.2017.37.5.739

3. Deng, Y. et al. (2022). Applications of knowledge graphs for food science and industry. Patterns, 3(5), 100484. DOI: 10.1016/j.patter.2022.100484

4. Haussmann, S. et al. (2019). FoodKG: A Semantics-Driven Knowledge Graph for Food Recommendation. International Semantic Web Conference (ISWC). http://www.cs.rpi.edu/~zaki/PaperDir/ISWC19.pdf

5. FoodOn Consortium. (2024). FoodOn: A farm to fork ontology. https://foodon.org/

6. Anthropic. (2025). Structured Outputs Documentation. https://docs.anthropic.com/en/docs/build-with-claude/structured-outputs

7. Anthropic. (2025). Claude Sonnet 4.5 Model Card. San Francisco: Anthropic.

8. Radford, A. et al. (2023). Robust Speech Recognition via Large-Scale Weak Supervision. Proceedings of the 40th International Conference on Machine Learning (ICML). https://cdn.openai.com/papers/whisper.pdf

9. Popovski, G. et al. (2024). Deep Learning Based Named Entity Recognition Models for Recipes. Proceedings of the 2024 Joint International Conference on Computational Linguistics (LREC-COLING). https://aclanthology.org/2024.lrec-main.406.pdf

10. Chia, Y.K. et al. (2022). Enhancing Food Ingredient Named-Entity Recognition with Recurrent Network-Based Ensemble (RNE) Model. Applied Sciences, 12(20), 10310. DOI: 10.3390/app122010310

11. Palermo, M. et al. (2014). A review of the impact of preparation and cooking on the nutritional quality of vegetables and legumes. International Journal of Gastronomy and Food Science, 3, 2-11. DOI: 10.1016/j.ijgfs.2013.11.001

12. Rinaldi, M. et al. (2022). Cooking at home to retain nutritional quality and minimise nutrient losses: A focus on vegetables, potatoes and pulses. Trends in Food Science & Technology, 126, 227-241. DOI: 10.1016/j.tifs.2022.10.009

13. Neo4j, Inc. (2024). Neo4j Graph Database Documentation. https://neo4j.com/docs/

14. Pinecone Systems, Inc. (2024). Pinecone Vector Database Documentation. https://docs.pinecone.io/

15. Levinson, C.A. et al. (2017). My Fitness Pal Calorie Tracker Usage in the Eating Disorders. Eating Behaviors, 27, 14-16. DOI: 10.1016/j.eatbeh.2017.08.003

16. Simpson, C.C. & Mazzeo, S.E. (2017). Calorie counting and fitness tracking technology: Associations with eating disorder symptomatology. Eating Behaviors, 26, 89-92. DOI: 10.1016/j.eatbeh.2017.02.002

17. Linardon, J. & Messer, M. (2019). My fitness pal usage in men: Associations with eating disorder symptoms and psychosocial impairment. Eating Behaviors, 33, 13-17. DOI: 10.1016/j.eatbeh.2019.02.003

18. Murphy, E.W., Criner, P.E., & Gray, B.C. (1975). Comparisons of methods for calculating retentions of nutrients in cooked foods. Journal of Agricultural and Food Chemistry, 23(6), 1153-1157. USDA ARS Publication.

19. Tian, J. et al. (2024). Core reference ontology for individualized exercise prescription (EXMO). Scientific Data, 11, 1319. DOI: 10.1038/s41597-024-04090-y

20. Transparency-One. (2024). Tracing the World's Food Supply from Farm to Fork with Neo4j. Neo4j Case Study. https://neo4j.com/case-studies/transparency-one/

21. OpenAI. (2024). Whisper Large v3 Model Card. Hugging Face. https://huggingface.co/openai/whisper-large-v3

22. Gupta, S. et al. (2024). Building FKG.in: A Knowledge Graph for Indian Food. Formal Ontology in Information Systems (FOIS 2024). arXiv:2409.00830

23. Bai, Y. et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv preprint arXiv:2212.08073. Anthropic.

24. Wang, W. et al. (2024). Nutrition-Related Knowledge Graph Neural Network for Food Recommendation. Foods, 13(13), 2111. DOI: 10.3390/foods13132111