docs/PRESENTATION_PROCESS.md

Presentation Process

This is the simple version you can use when presenting the project.

One-Sentence Summary

Our system helps Hong Kong elderly users check suspicious messages or screenshots by comparing them with real scraped fraud-news cases and giving clear safety advice.

The app supports Traditional Chinese, Simplified Chinese, and English. In the web app, the answer language follows the selected interface/output language. If an API client omits the language or sends auto, the backend detects the suspicious message language; Traditional Chinese remains the default for Hong Kong Cantonese-style messages.

The Process

1. Collect Fraud News

We scrape Hong Kong fraud-related news articles and store the case details.

Think of this as building a small local fraud case library.

2. Convert Cases Into Searchable Vectors

Each case is converted into a numerical representation called an embedding.

For presentation, you can say:

We turn each news case into a searchable fingerprint, so similar scam patterns can be found quickly.

3. User Enters A Suspicious Message

The user can:

• paste a suspicious SMS, WhatsApp message, or email
• upload a screenshot

If the user uploads a screenshot, Qwen OCR first reads the text from the image.

4. Search Similar Cases

The user message is also converted into an embedding.

The system searches the Chroma vector database and retrieves the most similar fraud cases.

Example:

• user message says parcel delivery fee
• system retrieves similar parcel-delivery scam cases
• user message says investment return
• system retrieves similar investment scam cases

5. Qwen Explains The Risk

Qwen receives:

• the user message
• the retrieved similar cases
• strict elderly-friendly safety rules

It then gives:

• a fraud score from 0-10
• What Happened
• Red Flags
• What To Do
• the answer in Traditional Chinese, Simplified Chinese, or English based on the selected output language, with auto detection only when no language is selected

6. Safety Advice For Hong Kong

The answer always reminds users that if they feel suspicious, they should call:

ADCC Anti-Deception Coordination Centre: 18222

Why This Is Useful For Elderly Users

• They do not need to understand technical terms.
• They can upload screenshots instead of typing everything.
• The answer is short and clear.
• It gives immediate actions, not just a warning.
• It uses Hong Kong-specific fraud cases and Hong Kong help resources.

Presentation Demonstration Script

1. Open the app.
2. Paste a suspicious message, for example a fake parcel fee message.
3. Click 檢查風險.
4. Show the score.
5. Show the similar cases retrieved from the vector database.
6. Explain the three safety sections:
   • What Happened
   • Red Flags
   • What To Do
7. Point out the ADCC 18222 reminder.

Technical Architecture

flowchart LR
    A["User message or screenshot"] --> B["Frontend"]
    B --> C["FastAPI backend"]
    C --> D["Qwen OCR if screenshot"]
    C --> E["Embedding model"]
    E --> F["Chroma vector database"]
    F --> G["Similar fraud cases"]
    G --> H["Qwen risk explanation"]
    H --> I["Elderly-friendly result"]

Key Point To Say

The AI does not answer from memory only. It first retrieves similar cases from our embedded fraud-news database, then uses those cases to explain the risk safely and simply.