Website · All Hoverfly models · Maven Central

By Hoverfly. On-device personal information detector for Kotlin Multiplatform: Android, iOS, macOS, JVM desktop, JavaScript and WebAssembly. It finds phone numbers, UPI IDs, Aadhaar, PAN, card and bank numbers, emails, names, addresses and more in a message, and hides them before you save, show or send it.

import io.github.rajumark.hoverfly.hideout.Hideout

Hideout().use { hideout ->
    hideout.hide("Call me on 98765 43210, my UPI is raju@okaxis")
    // "Call me on [PHONE], my UPI is [UPI]"
    hideout.hide("Mera naam Pooja Gupta hai, aadhar 4521 8736 1292")
    // "Mera naam [NAME] hai, aadhar [AADHAAR]"
}

• Made for Indian apps. Indian phone formats, Aadhaar (also masked XXXX XXXX 1234), PAN, UPI IDs, IFSC, voter ID, driving licence and vehicle numbers, Indian addresses with PIN codes, and Indian names, in English, Hinglish, Hindi and other Indian languages. European languages work too.
• 19 kinds of personal info, each with exact character offsets and a confidence score.
• Leaves ordinary numbers alone. Prices, times, order IDs, UPI transaction refs, PNRs and cities stay visible.
• No dependencies. Inference is plain Kotlin. There is no ONNX Runtime, TFLite, ML Kit or native code.
• Private and offline. The model ships inside the library on every platform. There is no network, no permission and no telemetry, so the text you are protecting never leaves the device.
• Small. 8 MB of int8 weights. Android (minSdk 21), iOS, macOS, JVM desktop, JavaScript and WebAssembly, from Kotlin and Java.

// build.gradle.kts: commonMain, or any platform source set
dependencies {
    implementation("io.github.rajumark:hideout:2.0.0")
}

It's on Maven Central, so no extra repository is needed. Gradle picks the right artifact for each platform:

The Android-only 1.x releases are on JitPack: com.github.rajumark:hideout:v1.x.

Upgrading from 1.x on Android: Hideout(context) still compiles in Kotlin (deprecated). The model no longer needs a Context, so switch to Hideout(). Java code must change new Hideout(context) to new Hideout().

The sample app on an emulator. Every result is computed on the device.

English	Hinglish	Hindi	Nothing to hide
"Call me on [PHONE], my UPI is [UPI]"	"Mera naam [NAME] hai, aadhar [AADHAAR]"	"मेरा नंबर [PHONE] है"	"Order #40512378 will arrive by Friday, costs Rs 24,999"

The KMP sample on each platform:

Android	iOS	Desktop	Web (Wasm)

val hideout = Hideout()       // loads the model: do it off the main thread, keep one instance

hideout.hide("Account no 50100234567812, IFSC HDFC0001234")
// "Account no [BANK_ACCOUNT], IFSC [IFSC]"

hideout.find("Call Rahul on 98765 43210")
// [Pii(type=NAME, start=5, end=10, text=Rahul, score=0.998),
//  Pii(type=PHONE, start=14, end=25, text=98765 43210, score=1.0)]

// only some kinds
hideout.hide("Call Rahul on 98765 43210", types = setOf(PiiType.PHONE))
// "Call Rahul on [PHONE]"

// your own replacement
hideout.hide("OTP is 482913", replacement = { "*".repeat(it.text.length) })
// "OTP is ******"

hideout.contains("thanks, see you on Monday")   // false

hideout.close()                       // frees the model's memory

find() and hide() are thread-safe. Long text is handled in overlapping windows, so every word gets context on both sides. Offsets are UTF-16 indices, the same as String.substring.

With coroutines:

val hideout = withContext(Dispatchers.Default) { Hideout() }
val safe = withContext(Dispatchers.Default) { hideout.hide(message) }

From Java:

try (Hideout hideout = new Hideout()) {
    String safe = hideout.hide("Call me on 98765 43210");
}

threshold (default 0.5) is how sure the model must be. Lower it to hide more (fewer leaks, more false alarms), raise it to hide less. SECRET covers passwords, OTPs, PINs and CVVs. ADDRESS covers street, house or flat, building and PIN code; a city on its own is not hidden. DOB is a date of birth; other dates are left alone.

Measured on hand-written chat messages in 14 languages that were written after the model was trained and never used to build or tune it, and on held-out test sets. "Hidden" = the whole item is covered; "left alone" = messages with no personal info where nothing was hidden.

Where it falls short: a name right before a relation word is sometimes missed ("Hemant bhai"); a bare number after "at" can be hidden as an address ("starts at 7"); Wi-Fi network names can be taken as passwords; a word next to a found item is sometimes swallowed into it ("flat B-1102 …"). Chinese, Japanese and Thai are not supported.

sample/ is a separate Gradle build that uses the published library, never the source. It resolves io.github.rajumark only from Maven Local, or from Maven Central with -PhideoutRepo=central. It has a Compose Multiplatform app for Android, desktop and iOS, and a web page built for both Kotlin/JS and Kotlin/Wasm.

./gradlew :hideout:publishToMavenLocal
cd sample
./gradlew :androidApp:installRelease
./gradlew :desktopApp:run
./gradlew :webApp:wasmJsBrowserDevelopmentRun     # or :webApp:jsBrowserDevelopmentRun
open iosApp/iosApp.xcodeproj                       # run the iosApp scheme on a simulator

hideout/                       the library
  src/commonMain/              public API (Hideout, Pii, PiiType) and the model in plain Kotlin
                               (internal/: Text (units + shapes), Featurizer, SentencePiece, Network, UnicodeTables)
  src/{jvm,android,apple,web}Main/   the only platform code: NFKC normalization, model loading, a thread-safe unit cache
  src/modelData/               hideout.bin (int8 weights) · spm_pieces.tsv (tokenizer)
  src/commonTest/              parity with the reference on 197 vectors, API, long text, latency; runs on every target
sample/                        demo apps using the published artifacts
scripts/GenTables.java         generates UnicodeTables.kt (character classes) so every platform agrees
docs/                          website (rajumark.github.io/hideout)

On JVM and Android the model ships as Java resources in the jar/AAR. Kotlin/Native and the web have no resources, so the build compiles it into the library (generateEmbeddedModel).

./gradlew :hideout:jvmTest
./gradlew :hideout:testAndroidHostTest
./gradlew :hideout:connectedAndroidDeviceTest              # on a connected device/emulator
./gradlew :hideout:iosSimulatorArm64Test
./gradlew :hideout:macosArm64Test
./gradlew :hideout:jsNodeTest :hideout:jsBrowserTest
./gradlew :hideout:wasmJsNodeTest :hideout:wasmJsBrowserTest

The parity tests require identical units, token, unit and shape ids, the same labels and the same hidden text as the reference implementation on all 197 vectors, on every target.

Text is cut into units at every letter / digit / symbol boundary, so a found item maps back to exact characters (raju@okaxis, gives raju@okaxis). Each unit is split into SentencePiece tokens; each token also carries a hash of its whole unit and its shape (case, script, number of digits, the symbol itself, whether a space comes before it). A small bidirectional transformer (4 layers, 7M parameters, int8 weights with one scale per row) labels every unit as the start or inside of one of 19 types, or nothing. Long text is cut into overlapping windows of 128 tokens.

See PUBLISHING.md.

Hideout is one of nine small on-device models by Hoverfly, all with the same install, the same free tier and nothing sent to a server.

See them all on the Hoverfly website.

Free for up to 10,000 monthly active devices. You don't need an API key, an account or a license file: add the dependency and ship. It works in commercial apps too, with no limit on how often each device runs it.

How devices are counted. A monthly active device is a device that runs Hideout at least once in a calendar month. The limit applies separately to each product, each platform (Android, iOS, web…) and each Hoverfly model. Once a product passes it, you have 30 days to get a commercial license. The library keeps working and never checks in with a server.

Not allowed under any tier (unless agreed in writing):

• selling or redistributing Hideout or its model on its own, or inside another SDK or library
• extracting, modifying, fine-tuning or retraining the model weights
• using the model or its outputs to train or distill another model
• reverse engineering the model or its file format
• offering it as a hosted API for others

Custom models. Hoverfly also designs and trains small, fast on-device models for your needs: privacy filters, moderation, classification, language detection, smart replies and more.

Contact for a commercial license or a custom model: raju348636@gmail.com or +91 63533 21951 (call or WhatsApp).

Full terms: Hoverfly Community License.