Alimentación Inteligente AI

An AI companion that learns a user’s cultural food habits and recommends small, realistic changes to improve everyday meals without calorie counting or restrictive diets.

Comment

Inspiration

Alimentación Inteligente did not begin as a hackathon idea. Its origin goes back to 1996, when a member of our family with an academic background in chemistry began an extensive, self-directed review of books and texts about food, digestion, food combinations, and eating practices.

Over more than thirty years, he compared ideas from different sources, applied practices in his own daily life, documented his observations, and gradually synthesized that knowledge into a practical methodology.

It was not created from one book or one isolated theory. It emerged from a long process of:

  • reading;
  • comparison;
  • personal observation;
  • refinement;
  • practical application.

Over time, the methodology was shared informally with relatives, friends, and people in the family’s social circle. Their experiences, questions, and difficulties revealed an important challenge: knowledge about food can be difficult to apply because people do not eat under controlled conditions.

Daily eating decisions depend on:

  • routines;
  • family traditions;
  • schedules;
  • budgets;
  • work responsibilities;
  • restaurant options;
  • familiar foods;
  • willingness and ability to make changes.

The aspiration therefore became larger than creating another food tracker. The goal was to make decades of accumulated knowledge easier to understand, practice, and share beyond a small family and social circle.

Generative AI made this possibility more tangible. A conversational system can first learn how a person has learned to eat, understand their routine, constraints, preferences, and possibilities for change, and then translate a complex methodology into small and understandable interactions.

Alimentación Inteligente AI is the first digital prototype of that vision: combining a methodology developed over decades with GPT-5.6 for contextual understanding and a deterministic engine for applying the rules implemented in the demo.

The experiences that inspired the methodology are personal and observational, not clinical evidence. The prototype does not claim to diagnose, treat, or universally define healthy eating.

What it does

Alimentación Inteligente AI transforms a complex food methodology into a personalized, understandable, and progressive daily experience.

The application begins by learning about the user’s:

  • familiar foods;
  • daily routine;
  • usual meal sources;
  • preferences;
  • constraints;
  • possibilities for change;
  • personal objective.

GPT-5.6 then generates a personalized first reading and a prudent Intelligent Eating Index as a starting point.

Alimentación Inteligente seeks to differentiate itself from applications centered primarily on indicators such as weight and calories. Instead, it tries to understand how each person learned to eat, how their routine works, and which changes are realistic within their circumstances.

From there, the user can:

  • organize the day through flexible eating moments such as breakfast, mid-morning, lunch, afternoon snack, and dinner;
  • register what they ate, the time of the intake, and the type of meal;
  • describe foods freely instead of depending exclusively on a fixed catalogue;
  • indicate their own perception of how compatible the meal was with the methodology;
  • request a guided review of the registered foods;
  • compare their self-assessment with the methodology result;
  • review whether the required interval from the previous intake was respected;
  • receive a small and practical adjustment for a future occasion;
  • view all registered intakes in a daily timeline;
  • edit or delete saved records;
  • close the day and update longitudinal progress.

The system separates two responsibilities clearly:

GPT-5.6 understands the person, interprets their routine, constraints, preferences, and possibilities for change.

A deterministic TypeScript engine applies the food-combination rules, intervals, daily outcomes, and index progression implemented in the demo.

This separation reduces the risk of the language model inventing methodological rules while preserving the value of natural and personalized conversation.

The index does not measure health, weight, calories, or clinical outcomes. It represents the degree of affinity between the user’s eating routine and the philosophy proposed by Alimentación Inteligente.

How we built it

Alimentación Inteligente AI was built as a full-stack web application using:

  • Next.js
  • React
  • TypeScript
  • OpenAI Responses API
  • GPT-5.6
  • Codex
  • Vercel
  • GitHub
  • Vitest
  • ESLint
  • Browser localStorage

GPT-5.6 is used as the contextual and conversational layer. It receives structured information from the onboarding process to:

  • create a personalized initial reading;
  • interpret the user’s routine;
  • identify practical areas of observation;
  • adapt the language to the user’s circumstances;
  • generate realistic questions;
  • propose contextual reflections without inventing methodological rules.

The prompts include explicit guardrails. GPT-5.6 is instructed not to diagnose, make clinical claims, invent food-compatibility rules, or present itself as the source of truth for the methodology.

The deterministic methodology engine is implemented separately in TypeScript. It handles:

  • food normalization;
  • aliases and spelling variants;
  • recognition of known and unknown foods;
  • extraction of several recognized foods from one natural-language phrase;
  • food-group classification;
  • expansion of compound preparations;
  • compatibility and incompatibility rules;
  • required time intervals between intakes;
  • comparison of self-assessment and guided evaluation;
  • compliant, non-compliant, and not-evaluable results;
  • daily closure calculations;
  • streaks and index progression.

For example, the phrase “Arepa con queso, mantequilla y sal” can identify Arepa and Queso, while preserving Mantequilla and Sal as components that have not yet been parametrized.

Each intake stores both its timestamp and an explicit local calendar date. This prevents an evening meal in Colombia from being moved into the following day when its timestamp is converted to UTC.

The application includes two clearly separated experiences:

  • a real onboarding flow, where progress starts from the index generated for the current user;
  • an explicit demonstration profile called Miguel, which shows how the index and streak could evolve over several weeks.

Codex was used throughout the development process to:

  • inspect the repository;
  • implement focused improvements;
  • write migration logic;
  • create regression tests;
  • resolve edge cases;
  • open pull requests;
  • document the development process.

Changes were reviewed through GitHub and validated with automated tests, linting, and production builds before being merged.

The production application is deployed publicly through Vercel and connected to the OpenAI API through protected environment variables.

Challenges we ran into

One of the main challenges was translating a methodology developed over decades into software rules without allowing the generative model to reinterpret or modify those rules.

In early iterations, the AI could generate suggestions that sounded reasonable but did not necessarily represent the methodology accurately.

We solved this by separating contextual intelligence from methodological evaluation:

  • GPT-5.6 personalizes the interaction.
  • The deterministic engine remains the source of truth for compatibility and interval calculations.

Another challenge involved free-text food registration. A user may write:

  • “papas”;
  • “papas a la francesa”;
  • “café negro”;
  • “huevos tibios”;
  • “arepa con queso, mantequilla y sal”;
  • the name of a preparation containing several components.

We created normalization, aliases, compound preparations, word-boundary recognition, recognized-food lists, and transparent handling for foods that are not yet parametrized.

Unknown terms are not silently discarded and are not assigned an invented classification.

Time zones created another important edge case. A dinner entered at night in Colombia could be stored in UTC as part of the next calendar day. This caused it to be incorrectly evaluated as the first intake of a new day.

We solved this by adding an explicit local date and making the interval engine, timeline, and daily closure use that local date.

We also discovered that real users were initially seeing the demonstration trajectory of Miguel. We introduced a persistent and explicit demo mode so that sample progress appears only when the demonstration profile is intentionally selected.

A further challenge was designing an index that does not react dramatically to one isolated meal. Individual intakes provide information, but the index evolves only through evaluable daily closures.

Reaching 100 requires sustained consistency over time rather than one perfect record.

Finally, we had to communicate uncertainty responsibly. The application distinguishes among:

  • the user’s own perception;
  • the guided review;
  • known foods;
  • unknown foods;
  • methodology-specific rules;
  • not-evaluable situations.

Accomplishments that we're proud of

We are proud that the project transforms more than thirty years of accumulated family knowledge into a functional and testable digital prototype.

Our main accomplishments include:

  • creating a Spanish-first onboarding experience;
  • generating personalized first readings with GPT-5.6;
  • integrating the OpenAI Responses API into a public production application;
  • separating generative AI from deterministic methodological evaluation;
  • allowing free-text food registration;
  • recognizing several known foods inside natural-language phrases;
  • preserving unknown foods transparently;
  • comparing the user’s perception with the methodology result;
  • calculating intervals using the correct local calendar date;
  • creating an editable and persistent daily timeline;
  • implementing daily closures, streaks, and longitudinal index progression;
  • creating a separate demonstration profile showing several weeks of progress;
  • adding migration support for previously stored records;
  • completing a regression suite with 78 passing tests;
  • validating the application through linting and production builds;
  • using Codex and GitHub pull requests as an auditable development workflow;
  • deploying a working application publicly through Vercel.

We are also proud that the application does not try to hide uncertainty. When a food or preparation has not yet been parametrized, the system says so clearly and preserves the information for future development.

What we learned

We learned that using a powerful language model responsibly requires defining what the model is allowed to decide and what must remain outside its authority.

GPT-5.6 is valuable for:

  • understanding context;
  • interpreting routines;
  • adapting communication;
  • asking useful questions;
  • producing personalized initial readings.

However, methodological rules should not depend on probabilistic generation.

Combining GPT-5.6 with a deterministic engine produced a system that is both conversational and auditable.

We also learned that daily context is not a cosmetic personalization layer. Food choices are influenced by:

  • childhood;
  • family traditions;
  • familiar dishes;
  • restaurant menus;
  • budgets;
  • work schedules;
  • social situations;
  • the user’s willingness and ability to change.

Another important lesson was that transparency improves the product. Terms such as not evaluable, unknown food, and methodology-specific rule are more responsible than presenting uncertain results as definitive answers.

We learned that longitudinal behavior is more meaningful than judging isolated meals. For this reason, the index changes through daily closures and sustained practice rather than immediately increasing or decreasing after every intake.

From a development perspective, Codex allowed us to move from a conceptual idea to a structured implementation through small and reviewable iterations.

Regression tests became especially important as local persistence, migrations, demo data, AI responses, and deterministic calculations began interacting with each other.

Most importantly, we learned that artificial intelligence can help make accumulated human knowledge more accessible without replacing the origin, history, or boundaries of that knowledge.

What's next for Alimentación Inteligente AI

The current version is a functional prototype and the first digital expression of a broader vision.

The next stage would begin with a formal review and documentation process for the bibliography that influenced the methodology.

This would create greater traceability among:

  • original sources;
  • interpretations;
  • implemented rules;
  • future revisions.

We also plan to expand the food and preparation catalogue through a controlled methodology-management system.

Instead of modifying application code for every addition, authorized reviewers could manage:

  • foods;
  • aliases;
  • preparations;
  • rules;
  • explanations;
  • required intervals.

Other possible developments include:

  • authenticated user accounts;
  • secure cloud persistence across devices;
  • longitudinal reports over months and years;
  • broader regional food profiles;
  • support for additional languages;
  • collaborative review by professionals from relevant disciplines;
  • personalized reminders and follow-up questions;
  • improved handling of portions and preparation methods;
  • improved interpretation of restaurant menus;
  • tools for describing or photographing meals;
  • anonymized learning from unrecognized foods and recurring questions;
  • accessibility improvements for older adults and users with limited digital experience.

A future version could also distinguish more clearly among personal experimentation, methodology education, and features that would require professional or clinical validation.

The long-term aspiration is to make the accumulated knowledge behind Alimentación Inteligente understandable and accessible to people around the world while preserving its history, methodological boundaries, and responsible positioning.

Built With

Share this project:

Updates