Inspiration

Millions of Indonesians take notes in Bahasa Indonesia — meeting minutes, lectures, articles — but summarizers are English-first and cloud-only. We wanted a summarizer that works in both languages, keeps short notes private on the device, and still handles long documents.

What it does

Ringkasku turns pasted or uploaded notes into structured summaries (title, summary, key points, action items) in Bahasa Indonesia or English. You pick the language, length (short/medium/detailed) and style (bullets, paragraph, study notes). Every summary shows exactly where it was computed, and every summary is saved to your personal history in Cloud Firestore so you can reopen it later. It is live at https://ringkasku.web.app.

How I built it — Firebase AI Logic Hybrid Inference

Hybrid model (InferenceMode.PREFER_ON_DEVICE, Gemini 2.5 Flash): short notes run on-device with Gemini Nano via Chrome's Prompt API when available, and fall back to cloud Gemini otherwise. initializeDeviceModel() runs on the user's click, with a download-progress bar.

Cloud model (InferenceMode.ONLY_IN_CLOUD): documents longer than the Remote Config threshold (default 6,000 chars) are routed to cloud Gemini.

Structured output: every summary is requested as JSON (responseMimeType: "application/json" + a response schema) — the same schema works on-device and in the cloud. The UI shows the real route: the SDK reports response.inferenceSource (on_device / in_cloud) and each summary is badged accordingly.

More Firebase: summaries are saved to Cloud Firestore (per-user, anonymous auth, locked-down security rules); Remote Config controls the default length, max input, cloud threshold and welcome message without a redeploy; App Check (reCAPTCHA Enterprise in production, debug provider on localhost) protects the AI endpoint.

Challenges I ran into

The on-device model is environment-sensitive: in our Linux VM, LanguageModel.availability() reports "unavailable", so the app transparently falls back to cloud — the hybrid design is exactly what saves the experience there. On-device inference itself has not been verified on real desktop Chrome hardware yet. The Firebase AI Logic docs describe initializeDeviceModel() and response.inferenceSource; I confirmed both exist in the shipped Firebase JS SDK v13.4.0 build before relying on them. Without a Firebase project in the build environment, I added a clearly labelled Demo mode (local extractive summarizer) so the whole UI flow is testable end to end — it is never presented as an AI summary. The live deployment at ringkasku.web.app runs the real Gemini path.

What's next

PDF upload and multi-document collections, then a shareable summary link.

Built With

  • app-check
  • cloud-firestore
  • firebase-ai-logic-(hybrid-inference)
  • firebase-js-sdk
  • html/css
  • javascript
  • remote-config
Share this project:

Updates

Submission history