Skrolz — Short reads, one story at a time

Swipe. Read. Go. Skrolz is a TikTok-style feed for short content: posts and stories, one full-screen at a time. Powered by Gemini 3 for AI posts, Study Buddy, and content moderation.

Comment

Inspiration

We wanted short reads to feel as effortless as scrolling—no endless feeds, no clutter. People love short-form video for discovery; we bet the same format could work for text. Skrolz was inspired by the question: What if you could swipe through bite-sized content the same way, with AI that helps you create and go deeper when you want? We combined a TikTok-style vertical feed with Google Gemini 3 to power creation, moderation, and engagement—so every story is focused, safe, and scroll-stopping.

What it does

Skrolz is a vertical, full-screen micro-reading app: one story at a time. Users swipe through short posts (≤280 characters) and multi-slide lessons in a personalized feed. For readers, we offer a For You feed, bookmarks, focus mode, and transparent “why you saw this” explanations. For creators, we offer:

Feature Description
AI Post Enter a topic (and optional tone/length); Gemini 3 generates 3 distinct post variants in seconds.
Study Buddy (Premium) For any topic or lesson, Gemini 3 returns 2 study tips, 1 actionable step, and 1 multiple-choice quiz—with structured JSON and high reasoning.
Content moderation Every post and lesson is checked by Gemini 3 for safety (harassment, hate speech, etc.) before going live.

The app is bilingual (English + Arabic, RTL) and runs on iOS, Android, Web (PWA), and desktop (Flutter).

How we built it

  • Frontend: Flutter (Dart) with Riverpod and go_router. We use Supabase for auth, real-time, and storage; RevenueCat for premium (Study Buddy).
  • Backend: Supabase (PostgreSQL, Edge Functions on Deno). Three Edge Functions call the Gemini 3 API:
Function Gemini 3 model Role
generate-post gemini-3-flash-preview Creates 3 post variants from a topic; we use thinkingLevel: "low" and temperature: 1.0 for creative, on-brand output.
study-buddy gemini-3-flash-preview Returns tips, action, and quiz in JSON; we use thinkingLevel: "high" and responseMimeType: "application/json" for reliable, structured responses.
moderate-content gemini-3-flash-preview Evaluates content against safety guidelines; we use explicit cached contents for system instructions and built-in safety settings.

We use retries with exponential backoff for rate limits and validate outputs (e.g. 280-char limit, JSON schema) before returning to the app. The Flutter app talks to these functions via Supabase client; feed ranking and recommendations use Supabase + materialized views (no Gemini there).

Challenges we ran into

  • Structured output: Getting Study Buddy to return valid JSON every time was tricky; we combined responseMimeType: "application/json" with clear prompts and client-side fallbacks (strip markdown, regex extract JSON) and schema checks.
  • Latency vs quality: We tuned thinkingLevel per use case—low for fast post generation and moderation, high for Study Buddy so the quiz and tips feel thoughtful.
  • Safety and UX: We needed moderation to run on both user-created and AI-generated content without blocking legitimate posts; we relied on Gemini 3’s safety settings and a rule-based fallback when the API is unavailable.

Accomplishments that we're proud of

  • Three distinct Gemini 3 integrations in one app: creative (AI Post), educational (Study Buddy), and safety (moderation)—each with appropriate reasoning levels and response formats.
  • Production-ready patterns: Caching for moderation prompts, retries/backoff, token logging, and strict validation so the app stays stable under load.
  • Full product experience: Bilingual (EN/AR), offline-capable feed cache, and a premium path (Study Buddy) that showcases Gemini 3’s reasoning for users who want to go deeper on content.

What we learned

  • Gemini 3’s thinking level and response MIME types let us optimize per use case (speed vs depth) and get reliable JSON for the Study Buddy feature.
  • Cached contents for system instructions in moderation cut down repeated tokens and made our Edge Function more efficient.
  • Combining a vertical, story-first UX with AI-powered creation and moderation made the app feel cohesive—users get “swipe to read” and “AI helps me create and go deeper.”

What's next for Skrolz

  • Richer recommendations: Use Gemini 3 to generate natural-language “why you saw this” explanations from user preferences and engagement signals.
  • Multimodal: Support image + text in posts and lessons and use Gemini 3’s multimodal APIs for moderation and optional alt-text or summaries.
  • Smarter Study Buddy: Let users ask follow-up questions or request more quizzes on the same topic, with conversation context kept in Gemini 3.
  • Launch: App Store and Google Play rollout, plus a public PWA so judges and users can try Skrolz without installing an app.

Built With

  • cached-network-image
  • connectivity
  • dart
  • deno
  • drift
  • easy-localization
  • flutter
  • go-router
  • google-fonts
  • google-gemini-3-api-(gemini-3-flash-preview)
  • image-picker
  • lottie
  • postgresql
  • revenuecat
  • riverpod
  • share-plus
  • sqlite
  • supabase
  • supabase-edge-functions
Share this project:

Updates