ClearMessage — AI Phishing Detection

Making online safety simple with AI-powered scam detection, clear explanations, and actionable guidance for everyone.

Comment

ClearMessage — A Strange Message. A Clearer Next Step.

Making online safety understandable for everyone.

Inspiration

Every day, people receive suspicious messages pretending to be banks, delivery companies, or trusted individuals.

These scams often create urgency, pressure people into making quick decisions, and target those who may be less familiar with technology.

We wanted to build something different from a traditional cybersecurity tool.

Something simple enough for our grandparents to use, but powerful enough to demonstrate real machine learning.

That's how ClearMessage was born.

Our mission is simple: help people understand suspicious messages and make safer decisions.

What It Does

ClearMessage is an AI-powered phishing awareness assistant that transforms confusing messages into clear explanations and practical guidance.

Users can simply paste a message or upload a screenshot.

ClearMessage then:

  1. Reads the message — Extracts text from screenshots using browser-based OCR.
  2. Analyzes it with AI — Uses our own trained machine-learning classifier to recognize ordinary messages, spam, and SMS phishing patterns.
  3. Highlights warning signs — Identifies supported suspicious phrases, including urgent payment requests and pressure tactics.
  4. Explains the result — Presents understandable explanations without complicated cybersecurity terminology.
  5. Suggests next steps — Helps users decide what to do if they have clicked a link, shared information, or received a suspicious request.

Built for Real People

ClearMessage was designed with accessibility in mind.

Our interface features large buttons, minimal text, intuitive navigation, and a clean, friendly design.

Additional features include:

  • SafeScan: Upload screenshots instead of manually copying messages.
  • Interactive Explanations: Explore highlighted warning signs.
  • Read Aloud: Listen to safety explanations.
  • Privacy Shield: Mask personal information before showing a message to someone else.
  • ActionGuide: Receive practical recommendations based on your situation.
  • One-Tap Demo: Experience ClearMessage without providing personal data.

No account is required.

How We Built It

Unlike applications that simply forward messages to a general-purpose AI chatbot, ClearMessage uses a machine-learning classifier trained by our team.

Our Machine Learning Model

We prepared a dataset containing 5,873 labeled messages:

  • 4,844 legitimate messages
  • 439 spam messages
  • 590 smishing messages

We reviewed dataset labels, resolved inconsistencies, and quarantined 98 conflicting records to improve data quality.

Our final classifier uses:

TF-IDF + Logistic Regression

We trained and evaluated different configurations before selecting the final model.

Model Performance

On our previously evaluated test set, the model achieved:

Macro-F1: 0.864

Smishing Recall: 84.1%

The model correctly identified 74 out of 88 smishing examples in that evaluation.

These are experimental results from a previously evaluated test set, not independent real-world validation.

Technology

Our application combines several technologies:

Frontend React, TypeScript, modern CSS, and accessible web components.

Backend Python and FastAPI.

Machine Learning Scikit-learn, TF-IDF, and Logistic Regression.

Computer Vision Tesseract.js for browser-based OCR.

Automation A pipeline connecting text extraction, ML classification, evidence-based detection rules, and personalized safety guidance.

Testing Automated Python, frontend, and cross-browser tests.

The system processes screenshots in the browser, while confirmed text is submitted to our analysis backend.

Suspicious links are never automatically opened.

Challenges We Faced

1. Building Our Own ML Pipeline

Training a classifier was only part of the challenge.

We also needed to understand the quality of our data, identify conflicting labels, prevent data leakage, and evaluate the model appropriately.

2. Making AI Understandable

Machine-learning predictions are not always easy to explain.

We combined the classifier with transparent, rule-based evidence detection so that ClearMessage could highlight actual warning signs instead of inventing explanations.

3. Designing for Accessibility

Our goal was to build something that people of different ages and technical backgrounds could understand.

We repeatedly simplified the interface, reduced unnecessary text, improved keyboard navigation, and tested responsive layouts.

4. Protecting Privacy

We designed screenshot processing to happen locally in the browser and avoided automatically opening suspicious links.

We also created Privacy Shield and privacy-safe assistance summaries to help users avoid sharing sensitive information.

What We're Proud Of

We're proud that ClearMessage is more than a phishing classifier.

It connects real machine learning with a practical user experience.

A person can upload a screenshot, review the extracted text, understand potential warning signs, and receive guidance — all through a simple interface.

We also built the project with accessibility, transparency, and privacy as central design principles.

What We Learned

We learned that building useful AI isn't just about achieving a good accuracy score.

It's also about making the results understandable, recognizing uncertainty, and designing technology around real human needs.

We gained hands-on experience with machine learning, dataset validation, frontend development, automated testing, computer vision, and accessible interface design.

Most importantly, we learned how to turn an ML experiment into a working application.

What's Next?

We want to continue developing ClearMessage beyond EVOHACKS.

Our next goals include independent evaluation on more diverse English and Czech messages, improving detection of emerging scams, testing accessibility with real users, and preparing a fully verified public deployment.

Our long-term vision is to make digital safety easier to understand for everyone.

Why AI for Automation?

ClearMessage automates the process of turning an unfamiliar message into understandable information and practical next steps.

Instead of requiring users to manually examine suspicious links, identify warning signs, and research what to do, ClearMessage brings those steps together in one guided workflow.

One message. One analysis. A clearer next step.

ClearMessage is a research prototype developed by Team Trampolin for EVOHACKS 2026. It provides informational guidance and cannot guarantee that a message is safe or fraudulent.

Built With

Share this project:

Updates

Submission history