Lily

The project was born because my girlfriend kills all her plants. No existing app worked for me, and it was the perfect excuse to build something ambitious on a stack I wanted to explore in depth.

The monorepo has 9 packages (Turborepo, Bun): an Effect RPC API, a React Native Expo app, a Next.js marketing site, a React admin dashboard, an MCP server, two PostgreSQL databases (Supabase + Hetzner/pgvector), and a native scanning module. The whole backend is written with Effect, using dependency injection via Context.Tag, typed errors, composable layers, and structured logging. No try/catch, no any. Every error is modeled in the type system.

The scanner identifies a plant from a photo (OpenAI Vision API) or reads a garden center label to pre-fill the care info automatically. Adding a plant takes a few seconds instead of manual research.

Each plant has its own AI assistant (GPT-4o-mini). The answers don't come from a generic prompt. The model gets the full context: species, health, last watering, care history. To go further, I built a full RAG pipeline: ingestion of Reddit discussions and specialist articles, chunking tuned per source, embeddings via text-embedding-3-small, storage in pgvector on a dedicated database, and similarity search before generation.

The MCP server exposes the user's data and actions to ChatGPT over OAuth and WebSocket. You can list your plants, check overdue care, or log a watering straight from a ChatGPT conversation, without opening the app.

The API covers 17 business domains: magic link auth, plant CRUD with care schedules, activity history, event-driven achievements, Expo push notifications, free/paid subscriptions through Stripe, and social features (follows, public profiles, handing off plant care while you're on vacation). It all runs on a Redis pub-sub event bus with exponential retry and a dead letter queue.