2026

AI Case Study

AyurvApp

Eat by your dosha.

Domain

Mobile-first web app, built with AI

Client

Personal project

Tools

Gemini, CodePen, Claude, Claude Design, Claude Code

What is Ayurveda?

Ayurveda is a traditional system of medicine that originated in India over 3,000 years ago and benefit physical, psychological, and spiritual health. In Ayurveda, every person is described through three energies called doshas. Each one is linked to natural elements and to a set of qualities. It focuses on whole-body healing and states that a person’s dosha determines their personality and health. Each food, with their specific characteristic can be favorable or not for each dosha.

🌬️ Vata is associated with air and movement, and is often described as light, quick and changeable.
🔥 Pitta is associated with fire and transformation, and is described as sharp, warm and driven.
⛰️ Kapha is associated with earth and water, and is described as steady, calm and grounded.

About the project

AyurvApp is an app that tells you in a few seconds whether a food is favorable for your Ayurvedic dosha (Vata, Pitta or Kapha), whether you are shopping at the supermarket or choosing a dish at a restaurant. It is a personal project, built to experiment with designing and building a product with AI, and it is still a work in progress: a prototype that will go online soon. I first built it with Gemini, because that was where I already kept my chat of food questions. I wanted something free, easy to access and not necessarily installed, and I was not yet using Claude Code, so I used Gemini to generate the code and CodePen to write and host it. A few months later I picked the project up again and moved it to Claude Code and Claude Design.

This case study follows that evolution, from phase 1 on Gemini to phase 2 on Claude: how I framed the problem, how I worked with each tool, which decisions were mine and which parts I delegated to AI, and how the product, the data and the design changed.

Important notice

AyurvApp is an experimental prototype built to explore a design process with AI. It is not a medical product and has no medical purpose. It does not diagnose, treat or advise, and it is not a substitute for a qualified Ayurvedic practitioner or a doctor. Food information comes from public sources, has not been verified, and may be incomplete or inconsistent.

A chat with Gemini became a tool.

I wanted to check, in seconds, whether a food suits my dosha while shopping or eating out. So I turned a long chat of food questions into a web app built on CodePen and saved as a shortcut on my iPhone, for my own use only.

Brief

Workflow

The brief for the first version was defined but loosely directed: build an app with three sections, a food search, recipes and a symptom log, that I could consult at the supermarket or in a restaurant. It had to be free, easy to put online, mobile first and without log-in, and it had to include foods I don’t usually eat. I set a loose goal, at least 20 foods, each marked as dangerous, caution or excellent, with an alternative for the dangerous ones. The rest, including the structure, the interface and many of the decisions, took shape along the way through the chat with Gemini.

Workflow

Step 1: the database. I generated a JSON database with Gemini. I specified the macro-categories I wanted (for example fruit, legumes, meat and vegetables) and the three markers used to classify each food: red, green and yellow.

Step 2: the interface. I asked Gemini to produce the HTML structure and the CSS, plus the JavaScript that reads the database.

Step 3: the data. I refined the data and integrated other JSON databases to make the final pool of foods more comprehensive.

Step 4: the UI. I refined the interface by giving Gemini change instructions for the CSS.

Step 5: extra features. Beyond food lookup, I added recipe suggestions and a tracker for my nasal symptoms.

POSITIVE OUTPUT
  1. Free and always at hand: it ran on CodePen as a shortcut on my iPhone, with no installation, no cost and no log-in.

  2. More than a lookup: besides food search it had recipes and a symptom log that kept its data after reopening.

  3. Personal fit: it covered foods I don’t usually eat, so I could check anything while shopping.

  4. Clear traffic-light verdicts: red, yellow and green markers made each food readable at a glance.

PAIN POINTS
  1. Loose brief: no defined scope, so features and structure grew along the way through the chat.

  2. Data I could not trace: the database was generated and merged from several files, with no clear source per verdict and no way to see where sources disagreed.

  3. No design system and interaction problems: the interface was refined with one-off CSS instructions, so it was harder to keep consistent.

  4. Accessibility not planned: no target level and no checks on contrast or touch targets.

  5. Private and fragile: it lived on a public pen with personal data in it, and could not be easily shared or put online as a product.

Gemini + Codepen

3

Main features

Claude Code redesign

I wanted to check, in seconds, whether a food suits my dosha while shopping or eating out. So I turned a long chat of food questions into a web app built on CodePen and saved as a shortcut on my iPhone, for my own use only.

Brief

Workflow

Brief


  • User and problem: someone who follows Ayurveda and needs to know, in seconds, whether a food suits their dosha, how to eat it, and what to choose instead. Someone who wants to have ayurbeda recipes.

  • In scope: choose a dosha and keep it; search one food at a time; a verdict on four levels with its source; raw vs cooked when they differ; at least one alternative; installable on the phone.

  • Out of scope: log-in, chatbot, payments, advertising, medical advice, diet planning, calorie counting, food combinations, mixed doshas.

  • Constraints: no server, no tracking, mobile first, installable web app, WCAG 2.2 AA as the target.

  • Data rules: one primary source, others only to cross-check; nothing invented; unverified or conflicting data always flagged; 40 foods in 9 categories.

  • Definition of done: ten checks that can be verified, for example the dosha and last food are still there after reopening the app.

Workflow

I worked with Claude in three roles, and I kept the same rule in all of them: I set the direction, it produced options, and I decided.

In the first role, Claude was a thinking partner. We wrote the brief together: I decided what was in scope and what was not, and Claude drafted and challenged it. It also compiled the food data and flagged where sources disagreed, while I chose the sources, the four verdict levels and the rule that nothing is invented.

In the second role, Claude Design was a production partner. Starting from my style tile, it generated tokens and components for the design system. I reviewed every output, corrected contrast, and rejected anything that did not fit the brand, such as the first wordmark.

In the third role, Claude Code was the builder. It worked from the brief and from written rules: a plan approved before any code, one feature at a time, and a decision log. After each step I checked the result in the browser and sent corrections, from the layout to the palette and the empty state.

The result is a process where AI does the fast production work and I keep the judgment: what to build, what to leave out, and when something is not good enough

POSITIVE OUTPUT
  1. A scalable system, not a one-off: a design system with tokens and components, built from my style tile and reused on every screen.

  2. Traceable data: 40 foods with a primary source, cross-checked with three others, and conflicts flagged on screen.

  3. A sharper flow: less actions, more direct results.

  4. Better UI quality: better UI output, better interaction.

PAIN POINTS
  1. Output less convenient than version 1: there is no always-on shortcut on my phone yet.install, offline use and deployment are still to do.

  2. Not yet validated: all 40 foods are flagged as unverified until I check them one by one, no real users have tested it.

  3. Weird user flow: despite on wireframe it all looked fine, it took me several attempts to adjust the user flow.

  4. Smaller scope: recipes are only a few examples and the symptom log is gone. A sharper brief meant giving up features.

Claude Code + Code Design

2

Main features