Your on-device AI job hunter.

Triage runs only when a notification arrives — typical inference is under 2 seconds on the GPU, and the Gemma model unloads after each pass. Background drain is dominated by the notification listener service itself (the same one any messaging app uses), not the LLM. Power profile measurements live in docs/architecture as we publish them.

No. The model loads into GPU memory on first triage call, runs inference, and unloads. Disk footprint is ~2.4 GB once (one-time download from HuggingFace) — RAM peaks during inference and returns to the OS afterwards. Implementation lives in MediaPipePromptRuntime.kt.

Only apps you explicitly enable in Settings. Banking and authenticator apps are not in the default list of 33 messaging/work apps — you'd have to opt them in. Even when enabled, content stays on-device because there's no cloud path: see network_security_config.xml (cleartext blocked, no upstream telemetry endpoint exists).

A standard HTTPS GET for one file (gemma-4-E2B-it.litertlm, ~2.4 GB) plus the User-Agent CueInbox-Android/<device-model>. No account, no token, no analytics ping. After the SHA-256 verifies (constant baked into the APK, see ModelDownloadManager.kt), the network listener can stop forever and the app keeps working offline.

Everything (notification history, model file, on-device knowledge graph) lives in the app's private storage. Uninstall wipes it. There is no cloud sync, no backup endpoint, nothing on a server with your name on it.

MediaPipe LiteRT's GPU delegate (the bit that lets Gemma run on your phone's GPU at usable speed) requires API 33 features. On older Android we'd have to fall back to CPU, which makes the same triage take 15–30 seconds — not a usable product. The minimum will lower as device drivers improve; tracked in ADR-008.

Soon — once Google's CASA security review for the notification-listener permission completes (see CASA runbook). Until then, the signed APK on Firebase App Distribution is the same build that will ship to Play. Drop your email below to get one ping when the listing goes live.

The full repo lives at github.com/avelikiy/cueinbox-android. Issues and pull requests are welcome — the threat model and ADRs in docs/architecture are the right place to start if you want to challenge a privacy claim.