Inspiration

Field researchers, journalists and teachers often work where the internet is unreliable — a mangrove boardwalk, a village classroom, a night market. FieldNotes lets them capture structured field notes offline and get instant AI summaries on-device, with the cloud as a fallback — never the other way round.

What it does

  • Structured field notes (title, tags, location, text) stored locally first, so the app works fully offline.
  • Instant on-device summaries: each note is summarised by Gemini Nano on-device via Firebase AI Logic hybrid inference — private, free, no connection needed.
  • Cloud fallback: if no on-device model is available, the same call falls back to gemini-3.5-flash-lite in the cloud.
  • Proof of where inference ran: every summary shows a badge — On-device, Cloud, or Demo mock — plus latency and model name, so the hybrid behaviour is visible, not just claimed.
  • Deep field reports: cross-note themes, action items and connections, analysed in the cloud (exactly 4 themes, 4 action items, 3 connections).
  • Firestore sync when back online, anonymous Auth, Remote Config (summary length, cloud analysis on/off, hybrid mode), App Check, search, tag filters, Markdown/JSON export. ### How Firebase AI Logic is used summarizeNote() uses InferenceMode.PREFER_ON_DEVICE: Gemini Nano in Chrome when available, otherwise the cloud model, with a JSON response schema (summary, tags, keyPoints). The returned inferenceSource (on_device | in_cloud) is stored on the note and rendered as a badge. analyzeFieldReport() uses a plain cloud model for cross-note reasoning — on-device inference only supports single-turn text generation, so long, multi-document reasoning belongs in the cloud. ### Challenges
  • gemini-2.5-flash was retired mid-build; gemini-3.8-flash took ~180s per call. Settled on the proven gemini-3.5-flash-lite.
  • Deep analysis kept returning empty action items and connections: the bug was in the code, not the prompt — ObjectSchema's 3rd constructor argument is optionalProperties, not required, so no field was required at all. Removing it (plus exact-count schema constraints) fixed it.
  • Summarise hung while Gemini Nano was still downloading: the warm-up call is now fire-and-forget, so the cloud fallback answers immediately. ### Accomplishments Real inference verified end to end on the live site: cloud summaries in ~2.3s with badges and model names, deep analysis 4/4/3 in ~2.4s, Firestore sync confirmed. The demo video shows the real app, not a mock. ### What's next
  • Show the on-device badge live once Gemini Nano finishes downloading.
  • Audio-note capture with on-device transcription.
  • Shared field projects for research teams.

Built With

  • firebase-ai-logic-(hybrid-inference-sdk)
  • firebase-app-check
  • firebase-remote-config
  • firestore
  • gemini-3.5-flash-lite
  • gemini-nano
  • react
  • typescript
  • vite
Share this project:

Updates

Submission history