Building Imi, an AI nutrition coach, from brand to React Native front end

Product
Imi · iPhone and Android
Timeline
Apr to Nov 2026
Role
Co-founder · design, code, marketing
Status
Beta with 20 testers · getimi.app
Team
Jashan Gupta, with Aman Arora (co-founder, AI infrastructure)
Imi on three iPhones: the planned week, Home, and the meal plan, on the brand's clay

02Background

It started with my own meal plans in ChatGPT.

ChatGPT was good at the plan and bad at the journey: last week's plan was buried in old chats, and nothing tracked my goal.

  1. A real problem

    General AI tools are too broad to carry a months-long nutrition goal.

  2. A design playground

    Push new design tools as far as they go, and ship the front end myself.

  3. A partner

    Aman Arora, an engineer friend who wanted to sharpen his AI infrastructure skills.

I wore five hats: brand, product design, product vision, the React Native front end and marketing. Aman built the AI infrastructure.

03Problem

Every nutrition app looked like a spreadsheet, and started after you'd eaten.

  1. Bright primaries or black

    Blue, orange, coral, lime, cyan. Nobody owned warm, earthy colour.

  2. Trackers, not planners

    Most count what you ate, after you ate it. Few say what to eat next.

  3. Built for people who already track

    Not for someone starting a GLP-1 or watching blood sugar, who needs a plan more than a log.

The category

04Product

Imi plans the week first, so you only log what changed.

Pick a goal from 10, and Home, your targets and the plan rearrange to match. Ask Imi to change the week and the change arrives as a card, with Undo.

Change the week by asking“I'm out Saturday night” clears that dinner, with Undo.
Snap a plate, get the numbersOne photo becomes 583 kcal and the macros.
Log a GLP-1 doseDose, time and injection site in a few taps.

AI-native planning

1 of 2

One Home per goal

Imi Home for someone on a GLP-1: protein 69 of 118 g, water 1.5 of 2.5 L, and the next dose on Tuesday
On a GLP-1Protein, water and the next dose lead.
Imi Home for blood sugar: calories, carbs by meal, added sugar and fiber, with glucose below
Blood sugarCarbs by meal sit next to glucose.
Imi Home for blood pressure: sodium against target, fiber, and a 128/82 reading
Blood pressureSodium and readings come first.
Imi Home in pregnancy: folate, iron, iodine and caffeine tracked as meals are logged
PregnancyFolate, iron, iodine and caffeine.

Widgets and Apple Watch

2 of 2

05AI workflow

Two people, 189 merged PRs: agents write, people gate.

Commits since 12 April
2,482
Co-authored with Claude
67%
Task worktrees
113
Tests in the October regression
6,826

I design in production. Claude Code writes most of the code inside rules I set, and every step that ships ends at a check or a person.

  1. Person

    Set the bar

    PRODUCT.md, a design skill and a 2,429-line voice guide load before any UI or copy.

  2. Agent

    Spec, then plan

    Claude drafts both from a brainstorm with me.

  3. Agent

    Build in a worktree

    Claude Code on one isolated branch per task.

  4. Script

    Checks

    Typecheck, lint and Jest on every PR, plus 6 Maestro end-to-end flows.

  5. Person

    Review and merge

    189 PRs merged into main.

One feature, end to end · 43 specs and 42 plans so far

I point agents at design quality too. In June, 35 agents audited the app against Nielsen's heuristics: 161 findings, 2 refuted, 22/40. A checklist where an item only counts with a commit and a screenshot took it to 32/40 in four gates.

The verdict

1 of 2

Baseline, 22 of 40

2 of 2

06Brand

Third name, one typeface, colour named for its job.

Imi was the third name, after Root and Esa, and 44 candidates fell to trademark, App Store or domain conflicts. Four typefaces became one.

Name

1 of 9

Logo mark

2 of 9

Finishing the mark

3 of 9

Typography

4 of 9

Finding the clay

5 of 9

No colour does two jobs

6 of 9

Nutrition colour

7 of 9

Iconography

8 of 9

Surfaces

9 of 9

07Illustration

218 paintings, so 39,774 recipes stopped sharing pictures.

73% of recipes showed a repeated or generic picture. 218 ingredient paintings now cover 99.5% of them, and the character grew up three times before it held.

Ingredients

1 of 3

The character

2 of 3

Where the character lives

3 of 3

08Website

I built getimi.app with a front door for every goal.

Goal hubs, from GLP-1 to PCOS
10
Sourced guides
67
Free tools and calculators
20
Commits since 27 September
104

Designed and built by me in Astro on Vercel: getimi.app. Each goal gets its own hub, guides and tools; scroll any page below, on desktop and on a phone.

getimi.app home page, full page on desktop
getimi.app home page, full page on a phone
Open getimi.app ↗

09Store

App Store and Google Play assets, rendered from code.

Store screenshots, iPhone and Android
132
Sets, one per goal
7
30-second app previews
2

A script frames simulator captures under each headline in headless Chrome and uploads them through the App Store Connect API, so a new headline is a one-line edit.

App Store screenshot 1
App Store screenshot 2
App Store screenshot 3
App Store screenshot 4
App Store screenshot 5
App Store screenshot 6
App Store screenshot 7
App Store screenshot 8
App Store screenshot 9

10Merch

The mark had to survive cotton, not just a screen.

Two tees, clay and white, carry the current mark and the site's line: "Don't just count calories. Plan them." They went to the printer as vector files; the cap, folded tee and mug are mockups from the same system.

Tees, cap and mug

11TikTok

Six agents plan the content week. None of them can post.

@imi.app on TikTok@imi.sundayreset
Posts in 30 days, 4 channels
185
Median views per TikTok
334
Best post: Chick-fil-A protein list
1,118
Carousels over videos on TikTok
~2×

Scheduled Claude agents draft, render and schedule; a person approves every post in a Claude artifact that works as a command center. Out of scope on purpose: auto-posting, paid ads and UGC creators.

  1. Data

    Metrics pull

    Postiz, App Store Connect and site counters, daily.

  2. Agent

    Scout and strategist

    Sunday trend scan; Monday brief with up to 5 proposals.

  3. Agent

    Drafter

    Up to 8 drafts a week, each claim-checked and tagged.

  4. Script

    Claim lint

    Only 11 approved claims pass.

  5. Person

    Approve

    Jashan or Aman. Agents never post, reply or change canon.

  6. Agent

    Render and schedule

    Higgsfield fills missing photos; Chrome and ffmpeg render slides; Postiz posts.

Daily and weekly · claims checked by a script, not a model
Chick-fil-A, small orders
Chick-fil-A, small ordersThe best post: 1,118 views.
Protein per 100 calories
Protein per 100 caloriesA ranked list, USDA source on the slide.
GLP-1 small plates
GLP-1 small platesEvery post ends on Link in bio.

  1. Useful beats funny

    Practical 'save this' protein carousels beat memes and recipe posts.

  2. Text lives in a safe band

    TikTok crops photo slides to about 3:4. The renderer keeps every word inside a measured band.

  3. New accounts read as spam

    A fresh Reddit account asking research questions got treated as spam. Reddit is hand-posted only.

12SEO

Each Monday a Claude Code run opens one pull request. Merging is the gate.

Protein pages built from USDA data
107
Guides, each with 2+ sources
67
Search impressions by early November, up from 0
1,453
Search click so far
1

The site is a software factory with one human step. The agent gets a week of search data and a budget of about 10 minutes of my review.

  1. Data

    Search data

    Search Console, Bing, autocomplete on 16 seeds, 9 rival sitemaps.

  2. Script

    Rank the work

    Near misses, low click-through, gaps and decay. Plain arithmetic.

  3. Agent

    Claude Code

    Up to 120 turns. Every new health fact needs a source fetched that run.

  4. Script

    Lint gate

    Fails dose advice, weight-loss promises, diet-culture words, fewer than 2 sources.

  5. Person

    Merge to deploy

    One PR, about 10 minutes to review. Merging ships it.

Every Monday at 13:00 UTC · GitHub Actions

  1. Pages must carry facts no other page has

    Each protein page ranks its food by protein per calorie across 536 USDA foods, to stay inside Google's scaled-content rules.

  2. A week with no sources ships nothing

    When the agent couldn't fetch the papers, its PR said so and changed no content.

  3. Search is slow

    Zero impressions in the first 28-day window, 1,453 by early November, 1 click. Too early to call.

13Results

A live beta and one brand everywhere. Distribution is the open problem.

Commits in seven months, two-thirds co-written with Claude
2,482
Recipes showing a repeated picture, after 218 ingredient paintings
73% → 0
Posts in 30 days, every one approved by a person
185

  1. In beta

    20 testers on TestFlight; testers get 3 months of Imi Pro.

  2. One system everywhere

    App, widgets, Watch, website, store pages, TikTok slides and a tee share one clay, one typeface and one mark.

  3. Distribution is early

    21 beta applications and 17 sign-ups in 28 days. The loops run; volume hasn't come yet.

14Learnings

Agents made the work cheap. Gates made it good.

  1. Design the gate, not just the agent

    Every pipeline here ends at a merge, an approval or a script that can say no. That's where my judgment sits.

  2. I'd start distribution with the first commit

    The app began in April and the site in late September. Search and content compound, and they started five months late.