01Consumer AI · iPhone

AI nutrition app

It answers one question on every screen: what fits in the rest of today?

Client: Our own product
Platform: iPhone and Apple Watch, over an edge back end
Stage: In development, launching on iPhone
Team: Our core team, from first sketch to store

Satchel, the app’s felt bear, checking a phone at the kitchen table beside a bowl of soup

The product

A food diary that looks ahead.

Most nutrition apps keep a careful record of the past. This one starts from the hours still to come: what fits in the rest of today, and what to eat next.

Logged the way it happened.A photo of the plate, the label or the menu, a sentence spoken or typed, or a barcode.
Honest about what it saw.An estimate arrives as parts with the portions it assumed, never one number to take on trust, and nothing counts until you tap Log.
Help with the next meal.What the cupboard can make, which of two dishes fits better, and a real recipe for tonight.
Calm by design.Food is never graded, exercise never earns food, and a missed day is a gap, not a zero. No advertising, and data is never sold.

On iPhone and Apple Watch, in widgets and on the Lock Screen, with Apple Health both ways.

Satchel at an open fridge, checking a phone

Inside the app

One question, on every screen.

Where the AI does its work: reading a plate, hearing a meal, finding the next one and correcting a log in plain words.

TodayOne question answered first: the calories and protein left in the day, then what has been logged.
PhotoA plate read into its parts, each with a portion you can change before anything is logged.
VoiceA meal said in one sentence, each dish picked out and held until you choose to add it up.
PantryWhat the food you keep can make, and what each dish would leave of today.
CompareTwo options side by side, and the rest of today with each.
CoachA meal corrected in plain words. The coach proposes the change; you accept it.

At a glance

The answer, without opening the app.

Widgets, the Lock Screen and Apple Watch carry the same reading as Today: what is left of the day, and what fits next.

The large widget: 1,140 kcal and 88 g of protein left, carbs, fat, fibre and sugar, and a meal that fits now — **Large widget**The whole day, and a meal that fits now.

The medium widget: 1,140 kcal left, 860 kcal logged, 88 g of protein left and every macro — **Medium widget**Calories and protein left, with every macro beside them.

The small widget: 1,140 kcal and 88 g of protein left — **Small widget**The two numbers that matter.

The app on Apple Watch: 1,140 kcal left, protein, carbs and fat, with water and recents — **Apple Watch**The same day on the wrist.

The app’s card on the iPhone Lock Screen: 1,140 kcal and 88 g of protein left — **Lock Screen**The day’s reading before you unlock.

How it’s built

Native, measured, one voice.

Built with our own method, from first sketch to App Store: a native iPhone and Apple Watch app over an edge back end, with AI that helps people log and understand what they eat, and never decides for them.

Principles

The person decides.The AI proposes; nothing enters the diary until the person taps Log.
Shows its working.Every estimate shows what it saw and where its numbers came from.
Measured, not assumed.Each AI feature is scored on held-out real-world sets before it ships.
Private by design.On-device recognition where it can be, attested devices, and deletion that deletes.
One voice.App, site, store listing and legal pages derive from the same typed sentences.

AI approaches

Multimodal understanding.Photos of plates, labels and menus, plus spoken and typed descriptions.
Grounded answers.Estimates anchored to reference nutrition data through retrieval, never free recall.
Structured outputs.Every response validated against a schema before the app shows it.
A model per task.Chosen by measured bake-offs on accuracy, speed and cost, and swappable.
Evaluation and cost control.Offline judging against truth sets, safety checks, and a ledger for every call.

Stack

Apple platforms: SwiftSwiftUIwatchOSWidgetKitLive ActivitiesApp IntentsHealthKitStoreKitVisionSpeechApp AttestSQLite, local-first
Back end: TypeScriptNext.jsEdge serverlessServerless PostgresObject storageZod
Quality: XCTestXCUITestStoreKit testingVitestAccessibility auditsMachine-enforced guardrails
Delivery: AI coding agentsDesign-first boardsContinuous evaluation

Case study coming soonThe full story of how it was built publishes when the app launches on iPhone.