Sathvik VK

Thaali

Photograph a thali, get every dish on the plate with its portion and calories, then correct anything the model got wrong.

Status
Launched. Live on the web as an installable PWA since Nov 2026. Not in the app stores
My role
Solo. Product, design, code and launch.
When
Nov 2026
Capabilities
LLM apps, Mobile, Web, Integrations
The Thaali home page with the headline Snap your thali and an illustration of a plate with several bowls.
thaali.life, captured 20 Nov 2026.Screenshot

The problem

Most calorie trackers are built around packaged food and single items. An Indian meal is several dishes on one plate, in portions nobody weighs. Typing each one in is the reason people stop tracking.

What I built

An Expo app that ships to the web as a PWA and can be added to an iPhone home screen.

  • Meal scanning: the photo is resized on the device, sent to an edge function, and a vision model returns the dishes and portions.
  • An onboarding flow that sets a calorie goal from the user's details.
  • Offline-first logging. Meals are stored in SQLite on the device and the app works without a connection.
  • Subscriptions through Razorpay on the web, with scan limits enforced on the server.
  • A food database of Indian dishes that I extended with regional entries and packaged products.

Decisions

The model only identifies foods and estimates portions. Calories and macros always come from the app's own food database, so a model cannot invent a nutrition number.

The nutrition logic is plain TypeScript with no React or Expo imports, and the same code runs in the app and in the server functions.

I chose Expo over a native iOS app to reach the web and iPhone from one codebase quickly. The cost is that store releases and device testing are still ahead.

How it fits together

Device
Camera
photo resized on device
Local log
SQLite, works offline
Server
Scan function
quota check, then vision model
Food database
the only source of nutrition numbers
Check
Editable result
the user corrects dishes and portions
A scan goes to the server once; everything after that is local. Simplified diagram drawn for this page, not a product screenshot.

Built with

ExpoReact NativeReact Native WebTypeScriptSQLiteZustandSupabasePostgresEdge functionsRazorpayJest
Next case studyOr press the right arrow keyB55A simulator for an AI-first handheld device: a Python agent daemon and a Qt QML interface, where every action that changes something needs approval.