Background
I wanted to actually learn how to build with AI, not just read about it. April 2025, I picked a problem I had: recipe websites are unusable. Ads, life stories, pop-ups. You just want ingredients and steps.
So I built a thing that takes a URL, runs it through Gemini, and gives you a clean recipe. That was Parsely v1. A Next.js app with one text input.
Where it went
It kept growing. I added user accounts, a cookbook to save recipes, folders, favorites, search. Then a streaming AI chat where you can ask questions about a recipe and it proposes edits you can accept or reject. Then subscriptions. Then an admin dashboard.
At some point I decided to go native. Built an iOS app in SwiftUI. Then Kotlin with Jetpack Compose for Android. Then a terminal CLI with Ink because I wanted to try building a TUI. Then a Chrome extension because sometimes you just want to parse from the toolbar.
Five platforms, one Convex backend syncing everything in real time.
Stack
| Layer | What |
|---|---|
| Web | Next.js 16, React 19, Tailwind CSS 4, shadcn/ui |
| Backend | Convex (real-time DB, auth, server functions) |
| AI | Gemini Flash Lite (web), gpt-4o-mini (CLI), Apple FoundationModels (iOS), ML Kit Gemini Nano (Android) |
| iOS | SwiftUI, Convex Mobile SDK |
| Android | Kotlin, Jetpack Compose, Hilt, Room, WorkManager |
| CLI | TypeScript, React 18, Ink v5, Puppeteer |
| Payments | Polar |
| Auth | Convex Auth (Google OAuth, magic links, password) |
| Analytics | PostHog |
| Hosting | Netlify (web), Convex (backend) |
What I learned
Every platform forced me to think about the same problems differently. State management in SwiftUI is nothing like React. Android's multi-module architecture with Hilt took a while to click. The CLI doesn't connect to the backend at all. It's a standalone scraper with Puppeteer and OpenAI, which was a deliberate choice to keep it simple and auth-free.
The AI parts taught me the most. Rate limiting that doesn't have race conditions. SSRF protection for user-submitted URLs. Structured output with Zod schemas. Streaming responses. On-device models on iOS and Android as fallbacks. iOS uses Apple's on-device FoundationModels when available, Android runs Gemini Nano through ML Kit, and the CLI just hits OpenAI directly since there's no user account to rate-limit against.