The problem

Pet cams are expensive, and plain noise meters fire on every slammed door or TV. Separation anxiety and barking complaints are hard to track when you are away from home.

The idea

PetGuardian AI reuses an old smartphone (plugged in at home) as a sentinel. One account, two modes in the same app:

- Monitor mode: the old phone listens and classifies sounds locally.

- Tutor mode: your daily phone gets push alerts, history and settings.

What it does

- Detects barks and howls and ignores TV, speech and urban noise.

- Alerts only when barking persists (configurable: N barks within M minutes).

- Plays a "calming voice" you recorded when the alert triggers.

- Daily history of barks and howls.

- Optional pre-event clip: the ~5 s before an alert, plus a text list of sounds heard just before it (doorbell, car, speech).

- Live listen-in (Pro), capped at 2 minutes per session.

- Free tier with ads; Pro via subscription.

How it works (tech details)

Audio pipeline, fully on-device

- Capture: PCM16 mono 16 kHz using the record package, sliced into 1 s windows.

- Model: YAMNet Lite (TFLite, 521 AudioSet classes) running in an IsolateInterpreter, so inference never blocks the UI thread. About 0.975 s (15,600 samples) per inference.

- Classification: uses per-class scores, not just top-1. Bark = Bark, Yip, Bow-wow at confidence ≥ 0.70. Howl, growl and whimper are tracked separately.

- Pre-filter: windows under a minimum dBFS level are discarded before inference.

- Alert logic: a sliding-window counter (default 5 barks in 3 min) with re-arming only after the count drops below threshold, so one long episode doesn't spam you. Parameters can be changed live without losing accumulated detections.

- Thermal control: an optional "analysis pause" (0–10 s) drops entire windows before inference. This cuts the real number of inferences instead of just delaying each one, which matters for an old phone that stays plugged in 24/7.

Alerts and backend

- Flutter app, Riverpod for auth state, Firebase Auth (email, Google, Apple).

- The Monitor never holds FCM credentials. It writes the event to Firestore and calls a callable Cloud Function notifyBarkEvent), which sends the push through the Admin SDK to the topic the Tutor subscribed to.

- Push text is localized (pt/en) server-side from the Tutor's saved language.

- Heartbeat every 45 s to show online/offline status.

- Subscriptions with RevenueCat; the entitlement is validated server-side in the Cloud Functions that gate Pro features.

Audio that leaves the device, and how

Analysis never uploads audio. Two opt-in/Pro features move audio, and both are peer to peer:

- Live listen-in: the Tutor creates a session through a callable function; the Monitor accepts it, and signaling goes through Firestore with ICE servers from another callable. Audio streams over WebRTC flutter_webrtc).

- Pre-event clip: a ~5 s ring buffer lives only in RAM. When an alert fires (and only if the user enabled it), it is written as WAV on the Monitor, kept for 3 days, then deleted. To listen, the Tutor Pro opens a WebRTC reliable DataChannel; the Monitor sends the file in 16 KB chunks with bufferedAmount flow control. The server only does signaling and never sees the audio.

Things I learned

- A reliable DataChannel can drop queued bytes if you close it too early, so the sender waits for bufferedAmount to reach 0.

- Delaying inference does not reduce heat; skipping windows does.

- Keeping the Cloud Functions keyless on the client side simplifies the security model a lot.

Privacy

Sound is analyzed on the device. By default nothing is stored. The pre-event clip is opt-in and off by default.

Stack

Flutter · Dart · TensorFlow Lite (YAMNet) · Firebase (Auth, Firestore, Cloud Functions, FCM, Storage, Crashlytics) · WebRTC · RevenueCat · AdMob · Supabase (app list) · gen-l10n (pt/en)

PetGuardian AI

PetGuardian AI

Comments

Publisher

Tech Stack

Sponsors

Outbid

Similar Products