Skip to content
Lakpriya S.
Case study 03 · My own AI apps

Four consumer AI apps, each live on iOS and Android.

Sparkle AI, MemeFrog, Tasklio AI and VisChef: chat, image, voice and vision features built with React Native and shipped to both stores.

  • React Native
  • OpenAI and other LLM APIs
Role
Developer and publisher
Team
Independent
Timeline
Jun 2023 – Jul 2025
Stack
React Native, OpenAI and other LLM APIs
4
AI apps
2
stores each
4
AI modes

The problem

Every AI feature looks easy in a demo. The hard part is making it fast, affordable and pleasant on a phone, keeping API keys off the device, and getting it through App Store and Google Play review.

Each of these apps uses a different AI pattern, and I picked the backend to suit it.

What I built

  1. 1. Sparkle AI: streaming chat (2023)

    A ChatGPT-style assistant in React Native, built in the months after the ChatGPT API launched.

    • OpenAI Chat Completions API with streaming over server-sent events (react-native-sse), so answers appear as they’re written.
    • GPT-3.5 Turbo or GPT-4, with custom instructions the user can edit.
    • Redux Toolkit with AsyncStorage for saved chats.
    • Native touches: voice input with the phone’s speech recognition, code blocks with syntax highlighting, and haptics while answers stream.

    I later turned the chat screen into an open-source library, react-native-sparkle-ai-chat, so other apps can drop it in.

  2. 2. MemeFrog: two models, one tap (2025)

    An Expo and React Native app that turns a topic into a finished meme.

    • A text model writes the joke as structured JSON: captions and a prompt for the picture.
    • An image model draws it in one of many visual styles, and the app adds the captions on top.
    • Moderation on what the user types and on what the models return, plus a way to report a meme.
    • Sharing with react-native-view-shot, and EAS Update for over-the-air fixes.

  3. 3. Tasklio AI: tasks by voice (2025)

    Tasklio AI lets people say a task out loud instead of typing it, then turns it into a structured task and helps prioritise the list.

  4. 4. VisChef: from a fridge photo to dinner (2025)

    I built VisChef in about a week for the Bolt.new hackathon.

    • Expo with Expo Router and Expo Camera.
    • Supabase: Auth, Postgres with row-level security, and Edge Functions that call the OpenAI API, so no key ships in the app.
    • GPT-4o mini vision spots the ingredients in a photo, then suggests recipes as structured JSON, shaped by the user’s diet and favourite cuisines.
    • Pexels API for recipe photos.

    VisChef grew out of Recipio, an earlier SwiftUI prototype of the same idea.

What carried across all four

Use structured JSON whenever the app has to act on what the model says, and always have a fallback. Keep keys on a server unless the user brings their own. Moderate what the model writes as well as what the user types. Stream anything longer than a sentence. All four apps are on the apps page, with store links.

See all apps