Inspiration
Long articles and lecture notes are easy to collect and hard to actually use. You paste a wall of text, lose focus after a few lines, and never go back to it. NoteWise exists to close that gap: paste a note once, and the app turns it into something you can actually study from — a short summary, a structured research brief, and a quiz that checks whether you understood it.
What it does
- Write or paste a note (a lecture, an article, meeting notes). Two sample notes ship with the app so it demos instantly.
- Summarize on-device — Gemini Nano, through Firebase AI Logic with InferenceMode.PREFER_ON_DEVICE, writes a 2–3 sentence summary and suggests a title and tags. Fast, private, works offline.
- Deep research — cloud Gemini turns the same note into a structured research brief: key ideas, background, connections, open questions.
- Quiz me — cloud Gemini writes 5 multiple-choice questions from the note. Answers are checked in the app, with a one-line explanation each.
Every result carries a badge stating exactly where it ran — On-device (Gemini Nano), Cloud (Gemini), or Demo mode — plus the latency of the call. If every live tier is busy, the app falls back through a multi-provider chain (Firebase AI Logic, /api/llm proxy, Pollinations, local demo) instead of showing an error.
How we built it
Vite + Firebase AI Logic, deployed on Vercel. The hybrid inference is the real integration in src/ai.js, imported and called at runtime. App Check is wired in (debug token in dev). The cloud model chain was verified live on 2026-11-13. The demo video was recorded as lossless PNG frames at 1920x1080 with real browser vector zoom so the UI text stays sharp.
Challenges we ran into
Gemini Nano is not on every device or Chrome version, so the on-device path could never be assumed — the whole app had to be designed around "on-device when available, cloud when not, demo mode when neither", with the badge telling the user the truth about where each answer came from. Google's error messages also name model versions that are already retired and return 404, so the cloud chain had to be re-verified model by model. Recording the demo was its own fight: smooth-scroll animations stutter under recording load, so the video uses instant jump cuts and state-verified clicks instead.
Accomplishments that we are proud of
A working hybrid-inference app where you can see, on every single answer, whether it ran on your device or in the cloud — and a fallback chain that keeps the app answering instead of erroring when a provider is busy. The demo video passed a frame-by-frame check at full zoom before it was approved.
What we learned
On-device AI is only useful if the app is honest about when it is actually on-device. The badge showing where each result ran turned out to be the feature users trust most, not a debug detail.
What is next for Notewise
Flashcards generated from any note for active recall, a mind-map view of the research brief, and export/share of notes with their summaries.
Built during the competition period: the whole app (notes + on-device summary + cloud research + quiz + fallback chain + demo video) was built Nov 9–14, 2026.
