Cozy World

What it is

Cozy World is a small shared space for two people, made to keep them connected day to day. The first screen counts how long it has been since the dates that matter to the couple, down to the hour, with a little hearts animation. Every day there is a new supportive message waiting. Longer notes live in categories, as cards you flip to read and then mark as read, so the unread ones are easy to find. A history screen lets you swipe through shared photos like a deck of cards.

Nobody has to ship an app update to add a message. The content lives in a database and is managed from a Telegram bot, so writing tomorrow’s message takes one chat conversation.

What I built

A small system of two parts, both open source under MIT.

The Flutter app (repository) is written fully by hand: zero AI-generated code. It shows what I can design and build on my own, without AI, from the architecture to the last widget. The core screens (counter, daily message, categories with flip cards, photo history) sit on a Clean Architecture base with data, domain and presentation layers. Android comes first, and the project builds for all Flutter platforms.

The Telegram bot (repository) is the content tool behind the app. Unlike the app, it was built with AI assistance under my direction. It runs as a Supabase Edge Function in TypeScript on Deno and works through inline buttons:

• Daily message: type the text, and the bot suggests the nearest free day; pick a type (text, video or audio) and confirm.
• Message in a category: pick a category or create a new one, then add the text.
• Edit and delete: paginated lists of daily or category messages, newest first, with a filter for unread categories.
• Cleanup: preview past daily messages, export them to the chat, then delete them.
• Date check: whether today and tomorrow already have a message, and how far ahead the calendar is filled.

Both parts share one Supabase Postgres database. The bot writes messages; the app reads them over GraphQL and writes back only the read flag.

Technical highlights

• Clean Architecture by hand. Use cases such as “get today’s message” and “change read status”, repository interfaces in the domain layer, and dependencies registered layer by layer: data, then domain, then presentation. Strings, dates and assets live in separate local packages split into an API and an implementation.
• GraphQL to Supabase through Ferry with generated, typed queries (build_runner, ferry_generator) against Supabase’s pg_graphql.
• Fresh data without over-fetching. A small “dirty state” mechanism marks category lists as stale after a read status changes, so only the affected lists reload.
• GetX for state and dependency injection, Navigator 2.0 (Router API) with my own route parser and delegate, including a 404 page and double back to exit.
• Config and secrets from --dart-define, with a fallback to flutter_secure_storage; local state in get_storage. CI checks every pull request and scans for leaked secrets.
• Bot as a state machine. Each chat’s flow and step are stored as JSON in a Postgres sessions table, so the serverless function stays stateless between Telegram updates.
• Secure webhook. The function accepts only POST requests that carry Telegram’s secret token header; everything else is rejected.
• Database does the bookkeeping. A trigger keeps the list of categories in sync with the messages table, so neither the app nor the bot has to maintain it.
• Delivery. The bot has Deno tests, and GitHub Actions runs them before deploying the function to Supabase.

Result

The app and the bot work together, and I keep adding to both: the app now also has a nutrition diary with a state journal. The screenshots come from a debug build in the iOS simulator. The app itself is Russian-only with a light theme; the English and dark versions were rendered from a local copy with translated strings and a dark theme built from the app’s own colours, only for these screenshots.

My role

Everything: the idea, the product design, the architecture and the app code, written by hand. The bot was a deliberate step onto an unfamiliar stack (TypeScript, Deno, serverless), built with AI assistance under my direction.