Google Gemma 4 Good Hackathon Android app Fully on-device

Gemma Guard is an Android app that spots phishing before you tap.

It captures the current screen, extracts visible text, and uses Gemma 4 locally to judge suspicious messages, login pages, and payment prompts. The result is a clear verdict, confidence score, reasons, and next-step guidance designed for everyday users and families.

Built for people who want a fast second opinion before tapping something risky.

  • Screenshot + OCR + Gemma 4
  • Private on-device analysis
  • Confidence, reasons, and action
  • Shareable with trusted contacts

Product demo

See the full flow: screen capture, OCR, Gemma reasoning, verdict, and user-friendly guidance.

Phishing succeeds because it shows up in ordinary phone moments.

Nadia is a 70-year-old mom. She uses her phone every day, pays bills, chats with family, and knows her way around apps. She still got caught by a phishing message that looked believable enough in the moment.

Gemma Guard was built for people like her: capable, busy, and deserving of a second opinion before a risky tap becomes a real loss.

A short, understandable flow for high-stress moments.

The app is designed to answer one question quickly: is this screen safe to trust right now?

  1. Capture the screen

    Grab the current Android screen exactly as the user sees it.

  2. Extract the visible text

    OCR pulls out the visible wording, links, brands, and urgency cues.

  3. Analyze image + text on-device

    Gemma 4 evaluates the screenshot and OCR text together on the device.

  4. Get a clear verdict

    Show verdict, confidence, reasons, and the safest next action to take.

Gemma 4 is central to the product, not just mentioned in the stack.

Gemma Guard uses Gemma because phishing screens are visual, contextual, and often written to pressure people into fast decisions.

Multimodal fit

The app needs a model that can reason over both the screenshot and the extracted text, not just one or the other.

Edge-friendly deployment

Gemma 4 is a strong fit for local-first Android experiences where privacy, latency, and offline behavior matter.

Plain-language output

The goal is not only classification. It is to turn model output into understandable reasons and a clear recommendation.

Built around the actual decisions users face on a phone.

Gemma Guard is not a generic warning banner. It gives context, explains risk, and keeps the core workflow local.

Works fully on-device

Screenshot analysis, OCR review, and reasoning happen locally with Gemma 4.

No cloud. No data leaves the phone.

Private messages and sensitive screens stay where they belong.

Image + OCR text together

Gemma Guard sees both the visual design and the written content.

Confidence score and reasons

Users get more than a warning. They get context they can understand.

Built for non-technical users

Short guidance, clear language, and simple next-step recommendations.

Share results with trusted contacts

Bring family into the loop when an extra opinion matters most.

Made for the kinds of screens that trick people in seconds.

Fake banking or payment prompts

Urgent money requests, payment confirmations, and bank warnings.

Account lock or password reset screens

Pages that push users to log in quickly before they stop to verify.

Delivery, refund, or package scams

Everyday messages that look harmless enough to tap without thinking.

Family safety check-ins

Results can be shared with a trusted contact when a second opinion helps.

The experience stays visual, readable, and calm.

These screens show the product as a real Android workflow, not a concept mockup.

Concrete implementation details that make the demo credible.

Screen capture + OCR pipeline

The app starts from the live screen instead of asking users to copy content manually.

Screenshot and OCR fed together

Gemma gets both the visual layout and the extracted text for better context.

Structured verdict output

The result is translated into verdict, confidence, reasons, and recommended action.

Local-first user experience

The core phishing check is designed to stay on-device for privacy and trust.

Designed for non-technical users

The hard part is not only detection. It is turning the answer into language people can act on.

A focused submission that shows a complete product loop, not just a model demo.

Gemma Guard was built for the Google Gemma 4 Good Hackathon as a privacy-first Android app that shows how Gemma can be used in a real mobile safety workflow: capture a risky screen, reason locally, explain the result, and help the user act safely.

  • Solves a real consumer phishing problem with a clear audience.
  • Uses Gemma 4 in a visible, product-critical way.
  • Connects model output to a simple Android UX for everyday users.
Read Kaggle write-up

Two builders, one specific person in mind.

Gemma Guard was built by Milana Kerbel and Pavel Kerbel for the Google Gemma 4 Good Hackathon. The mom at the top of this page is not a composite persona — she is Milana’s mother. She was scammed by a phishing message last year, and this is the app she should have had on her phone that day.

Contact: hello@gemmaguard.org

Fast answers for judges and first-time visitors.

What exactly does Gemma Guard do?

It captures the current Android screen, extracts visible text with OCR, and uses Gemma 4 locally to determine whether the content looks like phishing.

Why use Gemma instead of simple keyword rules?

Phishing is often about context and visual deception. Gemma helps the app reason over the layout, wording, and overall screen together.

Does the screen get uploaded anywhere?

The core experience is designed to run on-device, so private messages and sensitive screens do not need to leave the phone for analysis.

Who is this built for?

Everyday Android users, including older adults and families, who want a quick second opinion before tapping a suspicious screen.

Watch the demo and inspect the Android build.

The fastest way to evaluate Gemma Guard is to watch the product flow, then look at the Android implementation behind it.