Gatekeeper — on-device toxic message detection for Kotlin Multiplatform

Why Gatekeeper

Moderation that fits in your app

No server, no API key, no per-message bill. A small Kotlin library that you can read, test and ship.

Made for Indian chat

Understands native Hindi, Tamil, Telugu, Malayalam and Kannada, romanised Hinglish and Tanglish, and English, plus about 15 other languages. Catches misspellings like f*ck and chutiyaaa.

Few false alarms

Normal chat passes: 98.8% of everyday messages are not flagged. Slang like "this movie killed me" or "bhai tu toh kamaal hai" stays clean.

No dependencies

Pure Kotlin inference. There is no ONNX Runtime, TFLite, ML Kit or native code, and it runs on Android (minSdk 21), iOS, macOS, JVM desktop, JavaScript and WebAssembly, from Kotlin or Java.

Private & offline

The model ships inside the library on every platform. It makes no network calls, needs no permissions and sends no telemetry, so messages never leave the phone.

~1ms

Per message

Android emulator, once warm

~4MB

On-device

model + tokenizer

20+

Languages

native and romanised Indic

Dependencies

no native code

On device

Type a message, see the verdict

Real output from the sample app on an Android emulator. These are not mockups.

Gatekeeper says a friendly Hinglish message looks fine

Slang passes

"bhai tu toh kamaal hai 🔥" → Looks fine

Threat caught

"tujhe jaan se maar dunga" → Flagged · Threat

Insult caught

"you are a stupid idiot" → Flagged · Insult

Same library on every platform

The Kotlin Multiplatform sample app, built from the published artifacts, gives the same result on Android, iOS, desktop and the web.

Android

you are a stupid idiot → Flagged · Insult 99%

iOS

you are a stupid idiot → Flagged · Insult 99%

Desktop (JVM)

you are a stupid idiot → Flagged · Insult 99%

Web (Wasm)

you are a stupid idiot → Flagged · Insult 99%

Install

Add it to your build

One line in your Gradle file, from Maven Central. Works in commonMain for Android, iOS, macOS, JVM desktop, JavaScript and WebAssembly. No API key, no account.

Maven Central live

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

// Gradle picks the right artifact per platform:
// gatekeeper-android · gatekeeper-jvm · gatekeeper-iosarm64 · gatekeeper-iossimulatorarm64
// gatekeeper-macosarm64 · gatekeeper-js · gatekeeper-wasm-js

Build with a prompt

Add Gatekeeper from rajumark to this Kotlin Multiplatform or Android project.

What it does: on-device toxic message detection. Given a chat message, comment or review, it says whether it is toxic (abuse, hate, threats, sexual harassment), with category hints. English, Hindi, Tamil, Telugu, Malayalam, Kannada, Hinglish and more. ~4 MB, no dependencies, no network.

Install (Maven Central, no extra repository):
  build.gradle.kts (commonMain or any source set): implementation("io.github.rajumark:gatekeeper:2.0.0")
  Platforms: Android, iOS, macOS, JVM desktop, JavaScript, WebAssembly

Usage:
  val gatekeeper = Gatekeeper()      // off the main thread, keep one instance
  if (gatekeeper.isToxic(message)) { /* hide it, ask to rephrase, or send to review */ }

Repo: https://github.com/rajumark/gatekeeper#readme
Add the dependency, then follow the README for the exact API and current version.

Usage

One call per message

Load once and reuse the instance. check() is thread-safe.

Kotlin

val gatekeeper = Gatekeeper()

val v = gatekeeper.check("tujhe jaan se maar dunga")
v.isToxic       // true
v.score         // 0.98
v.topCategory   // THREAT

gatekeeper.isToxic("bhai tu toh kamaal hai")  // false
gatekeeper.check(text, Sensitivity.STRICT)     // catch more

gatekeeper.close()

Java

try (Gatekeeper gk = new Gatekeeper()) {
    Verdict v = gk.check("you are a stupid idiot");
    if (v.isToxic()) {
        // hide, blur, or ask to rephrase
    }
}

Use cases

Where Gatekeeper fits

Anywhere people write to each other, and before anything goes to a server.

Chat & community apps

Blur abusive incoming messages behind a "tap to show", or nudge the sender before they hit send.

"chup kar bsdk" → hidden

Comments & reviews

Hold toxic comments for review instead of publishing them, with no moderation API bill.

"what a piece of shit app" → review queue

Dating & social apps

Catch sexual harassment and threats in DMs before they reach the other person.

"send nudes" → Flagged · Sexual

Kids & education apps

Use Sensitivity.STRICT in class chats and student forums.

"tu pagal hai kya be" → Flagged

Games & live chat

Fast enough to check every message in a busy lobby on the player's own phone.

"gg noob uninstall" → fine

A first filter before the server

Cut what reaches your cloud moderation, and keep working offline.

98.8% of normal chat passes locally

App ideas

20 apps you could build with Gatekeeper

Real problems Gatekeeper solves today. Pick one and ship it.

Chat apps · Blur abusive incoming messages behind "tap to show".

Think-before-you-send · Nudge users with "Are you sure?" when their own message is toxic.

Comments & reviews · Hold toxic comments for review instead of publishing them instantly.

Dating DMs · Catch sexual harassment and threats before they reach the other person.

Kids & edtech apps · Keep class chats and doubt forums clean with strict mode.

Game lobby chat · Fast enough to check every message in a busy lobby, on the player's phone.

Live stream chat · Hide toxic messages in fast-moving live chat without server lag.

Cheaper cloud moderation · Filter on the device first and cut your moderation API bill.

Marketplace chat · Flag threatening or abusive buyers and sellers early.

Rider–driver chat · Spot threats and harassment between riders and drivers.

Community & society groups · Keep neighbourhood and housing-society groups civil.

Workplace messaging · Flag abusive messages in team chat before they escalate.

Username & bio checks · Block abusive usernames, bios and group names at sign-up.

Chatbot input guard · Detect abuse aimed at your bot and reply calmly or end the chat.

News app comments · Keep comment sections under news articles readable.

Voice room transcripts · Check speech-to-text transcripts of audio rooms for abuse.

Live Q&A & polls · Keep toxic questions off the screen at events and webinars.

Review platforms · Allow harsh 1-star reviews, but not slurs or threats.

Parental safety · Warn a child on their own phone when an incoming message is abusive.

Offline moderation · Keep moderating when the network drops or the server is down.

Quality

Honest numbers

Measured on held-out test sets and on hand-written chat messages. Every model gets its own tuned threshold.

	Gatekeeper	Multilingual BERT toxicity classifier
Indian-language comments (hi, ta, te, ml, kn), F1	0.84	0.63
Same comments typed in Latin letters, F1	0.83	0.65
English comments, F1	0.63	0.50
15 world languages, F1	0.77	0.91
Hand-written chat messages, accuracy	88%	80%
Everyday chat not flagged	98.8%	85.1%
Hate-speech stress test (no slurs), accuracy	45%	61%
Size	3.8 MB	711 MB
Latency, one message, 1 CPU thread (laptop)	~0.1–0.4 ms	~16 ms

Gatekeeper wins on Indian languages, English, everyday chat and size/speed by a wide margin. The multilingual BERT is better on its own 15-language benchmark and on subtle group hate without slurs ("X are vermin"), which is Gatekeeper's weakest area; pair it with a server check if that matters for your app. Category scores (insult, profanity, threat, hate, sexual) are most reliable for English; decide on isToxic / score.

Reference

API at a glance

Member	Description
`Gatekeeper()`	Loads the bundled model. Implements `AutoCloseable`.
`check(text, sensitivity = BALANCED)`	Returns a `Verdict`. A blank message is never toxic.
`check(text, threshold)`	Same, with your own threshold (0..1).
`isToxic(text, sensitivity)` · `score(text)`	Shortcuts for the verdict and the raw score.
`Verdict(isToxic, score, categories)`	Plus `topCategory`: the strongest category when toxic.
`Category`	`INSULT`, `PROFANITY`, `THREAT`, `HATE`, `SEXUAL`
`Sensitivity`	`STRICT` (0.25) catches about 85% · `BALANCED` (0.42) · `RELAXED` (0.70) flags about 0.6% of normal chat

Pricing & license

Free until you're big.

No API key, no sign-up, no license file. Add Gatekeeper to your project and ship. It stays free until your product passes 10,000 monthly active devices, and there's never a limit on how often each device runs it.

Start here

Community

Free

For indie developers, startups and growing apps, up to 10,000 monthly active devices.

Commercial use allowed
No API key, account or license file
Unlimited calls on every device
Runs fully on-device, with no tracking or phoning home

Get started

Commercial

Let's talk

For products above 10,000 monthly active devices on any platform.

One license per product, per model
Pricing that fits your scale
Direct support from the people who built it
Early access to model updates

Email WhatsApp Call

Custom models

Built for you

A small, fast on-device model trained for your language, domain or task.

Moderation, classification, detection, replies and more
Tuned on your use case and your users' languages
Shipped as a plain Kotlin library, like Gatekeeper
Private: your model runs on your users' devices

Email WhatsApp Call

How devices are counted

A monthly active device is a phone, tablet or other device that runs Gatekeeper at least once in a calendar month.

The 10,000 limit applies separately to each product, each platform (Android, iOS, web…) and each Hoverfly model. Crossed it? You have 30 days to get a commercial license.

What the license does not allow

Selling or redistributing Gatekeeper or its model on its own, or inside another SDK
Extracting, modifying or fine-tuning 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

Questions about licensing? Email raju348636@gmail.com or call / WhatsApp +91 63533 21951. Full terms: Hoverfly Community License.

FAQ

Questions

What is Gatekeeper?

A Kotlin Multiplatform library (Android, iOS, macOS, JVM desktop, JavaScript, WebAssembly) that reads a message and tells you whether it is toxic: abusive, offensive, hateful, threatening or sexual harassment.

Which languages does it support?

Best on English, Hindi, Tamil, Telugu, Malayalam and Kannada, in native script or typed in Latin letters (Hinglish, Tanglish). It also works in about 15 other languages including Spanish, French, German, Russian, Arabic, Japanese and Chinese, less reliably. To detect the language first, use Beacon.

Which sensitivity should I use?

BALANCED for most apps. STRICT for kids' apps or when a person reviews what gets flagged. RELAXED when you hide messages automatically and want the fewest false alarms.

Does it run on device?

Yes. Inference runs entirely on the device in plain Kotlin. It makes no server calls, needs no network permission and sends no telemetry.

Where does it fall short?

Subtle hate against a group without slurs, sarcasm, and some English idioms ("I hate mondays"). Treat it as a strong first filter, not a final judge.

Is the Kotlin port exact?

Features and scores match the reference implementation on all 140 test vectors (max difference 5e-7), both on the JVM and on an Android device.

How much does Gatekeeper cost?

Nothing for products with up to 10,000 monthly active devices per platform, including commercial apps. There is no API key or account, and no limit on how often it runs. Above that, a commercial license is needed. See Pricing.

What happens when my app passes 10,000 monthly active devices?

Nothing breaks: Gatekeeper keeps working and never checks in with a server. You have 30 days to get a commercial license. Email raju348636@gmail.com or call / WhatsApp +91 63533 21951.

Can you build a model for my app?

Yes. Hoverfly trains custom on-device models for your language, domain or task and ships them as a small library like this one. Get in touch on the same email or number.

Try the real model, right here

Moderation that fits in your app

Made for Indian chat

Few false alarms

No dependencies

Private & offline

Type a message, see the verdict

Slang passes

Threat caught

Insult caught

Same library on every platform

Android

iOS

Desktop (JVM)

Web (Wasm)

Add it to your build

One call per message

Where Gatekeeper fits

Chat & community apps

Comments & reviews

Dating & social apps

Kids & education apps

Games & live chat

A first filter before the server

20 apps you could build with Gatekeeper

Honest numbers

API at a glance

Free until you're big.

How devices are counted

What the license does not allow

Questions

What is Gatekeeper?

Which languages does it support?

Which sensitivity should I use?

Does it run on device?

Where does it fall short?

Is the Kotlin port exact?

How much does Gatekeeper cost?

What happens when my app passes 10,000 monthly active devices?

Can you build a model for my app?

Gatekeeper is one of eight.