On-device · nothing leaves your phone

Notifilter

Blocks promotional Android notifications and keeps the ones that matter. A 19 MB model runs entirely on your phone — no account, no server, no notification ever uploaded.

Download

Android APK v1.0 · min Android 8.0

arm64-v8a RECOMMENDEDEvery phone sold in India since ~2018 18.9 MB armeabi-v7aOlder 32-bit devices 13.8 MB

Install the APK — Android will ask you to allow installs from your browser.
Open Notifilter → Grant notification access. Android only permits this from Settings.
Wait for the one-time 21 MB model download.
Watch History. When the decisions look right, turn off Dry run.

Measured

0.996KEEP-recall on held-out test

0.910Promotional F1

18 msPer notification, 1 thread

32 MBint8 model, downloaded once

The distinction it makes

BLOCK

Get ₹200 cashback on your next orderAn offer. You lose nothing by never seeing it.

KEEP

₹200 cashback credited to your walletA record of your money moving. Same words, opposite meaning.

A keyword filter cannot separate these. That pair, in twelve scripts, is most of what the model was trained on.

How a notification is decided

notification arrives
   ├─ on your allowlist?      → KEEP   never classified
   ├─ OTP / debit / delivery? → KEEP   a rule, not a probability
   └─ otherwise               → model  p ≥ 0.9724 → BLOCK

An OTP never reaches the model. The rule layer fires on 3.8% of traffic at 99.2% precision — a classifier gives you an average, and a one-time password needs a guarantee.

Languages

Script	Test rows	KEEP-recall	Promo F1
Latin / Hinglish	2436	0.997	0.911
Devanagari	178	0.990	0.946
Bengali	260	0.995	0.864
Telugu	23	1.000	1.000
Gurmukhi	22	1.000	0.769
Kannada	20	0.909	0.947

Tamil, Malayalam, Gujarati, Odia, Assamese and Urdu are trained but have too few test rows to quote a number honestly.

Model delivery

The APK ships without the model. Cloudflare Pages caps files at 25 MiB, so the 32 MB model is gzipped to 20.9 MB and split in two, each part SHA-256 verified on arrival and again after reassembly.

GET /models/manifest.json
GET /models/model.onnx.gz.000   10.5 MB
GET /models/model.onnx.gz.001   10.5 MB
     → concat → gunzip → verify sha256 → cache on device

Honest limits

Android gives a notification listener no pre-display veto. Filtering means cancelling right after post, so a brief flash is possible.
Trained on synthetic data plus public datasets — no real device capture. Expect drift against your actual notifications.
Blocked notifications are recorded in full and stay readable in History. Nothing is deleted.
Ships in dry-run mode: it classifies but removes nothing until you turn that off.
Feedback you submit is uploaded on purpose. Device stats are opt-in and off by default. No location permission is ever requested.