ScamShield

Stop Scams Before They Reach You

Comment

💡 Inspiration

Most of us don’t worry about ourselves when scams come up.
We worry about someone else.

An aging parent.
A non-tech-savvy relative.
Someone we love who might panic before thinking.

The inspiration for ScamShield came from a simple but uncomfortable question:

“If this message landed on my mum’s phone, would she stop and think or would she just react?”

This question became urgent with the rise of Social Engineering scams, especially the infamous "Hi Mum" text fraud.

We realized something critical: traditional spam filters are stuck in the past. They look for bad words like Lottery, Prince, or Bank. But modern scammers don’t use bad words — they use bad stories.

They exploit human psychology: fear, urgency, authority, and trust.

They don’t attack systems first.
They attack people.

We wanted to build a system that doesn’t just read text, but actually understands context — effectively giving every vulnerable user a rational second opinion at the exact moment emotion takes over, before a dangerous click is made.

⚙️ What it does

ScamShield is a Hybrid AI Cybersecurity Analyst. It allows users to scan text messages, emails, or chat screenshots to instantly assess their danger level.

Unlike standard filters, it uses a "Smart Override" Protocol defined by our Hybrid Intelligence logic:

$$ Risk_{final} = \begin{cases} High & \text{if } P(Spam){Bayes} > 0.8 \ High & \text{if } Intent{LLM} \in {Urgency, Finance, Auth} \ Low & \text{otherwise} \end{cases} $$

  1. Speed Scan:
    A lightweight Machine Learning model instantly checks for known spam patterns.

  2. Reasoning Scan:
    If the message passes the first check but feels wrong, our Llama 3 engine analyzes the intent behind the message.

  3. The Result:
    If psychological manipulation is detected (e.g., fake emergencies or impersonation), ScamShield overrides the “Safe” score to High Risk and explains why in clear, plain English.

The goal is not just detection — but understanding.

🛠️ How we built it

We built ScamShield using a Dual-Brain Architecture to balance speed and intelligence:

  • The Brain (Backend):
    We used Python (Flask).
    For the Speed Brain, we trained a Naive Bayes Classifier using scikit-learn, based on probability theory:

$$ P(Spam | w_1, \dots, w_n) \propto P(Spam) \prod_{i=1}^{n} P(w_i | Spam) $$

For the Reasoning Brain, we integrated Groq’s API running Llama 3 (70B), achieving LLM reasoning speeds under $2\text{s}$.

  • The Body (Frontend):
    Built with React, Vite, and Tailwind CSS, using Shadcn UI to create a clean and trustworthy interface.

  • The Eyes (OCR):
    An image processing pipeline extracts text from screenshots so users can scan WhatsApp chats directly.

  • Deployment:
    A split-host strategy using Render (backend) and Vercel (frontend).

😤 Challenges we ran into

The biggest challenge was "Dependency Hell."

  • The "It Works on My Machine" Problem:
    Our local environment was Windows, while production ran on Linux. pip freeze captured Windows-specific binaries (e.g., faiss-cpu) that crashed Linux builds. We manually rebuilt and sanitized requirements.txt to ensure cross-platform compatibility.

  • The "Cold Start" Problem:
    Hosting ML models on free-tier servers caused high latency. We solved this with our hybrid approach — letting Naive Bayes handle simple cases instantly, and invoking the LLM only when deep context analysis was required.

  • CORS & Integration Issues:
    Connecting a Vercel frontend to a Render backend introduced strict CORS errors, resolved through proper Flask header configuration.

🏅 Accomplishments that we're proud of

  • Defeating Contextual Fraud:
    Successfully detects "Hi Mum" scams by analyzing semantic intent instead of keywords.

  • Real-Time Speed:
    Groq-powered inference enables near-instant deep reasoning without sacrificing UX.

  • Clean Architecture:
    ML logic is fully decoupled from the API layer, making the system scalable and maintainable.

🧠 What we learned

  • Less is More:
    A smaller, well-prompted 8B model outperformed a 70B model for this task.

  • Deployment is a Skill:
    Writing code is only half the battle — shipping reliable systems requires DevOps discipline.

🚀 What's next for ScamShield AI

  • Browser Extension:
    Real-time Gmail and web content scanning.

  • WhatsApp Bot:
    Users can forward suspicious messages directly to ScamShield for analysis.

  • Audio Analysis:
    Detecting deepfake voice scams during live phone calls.

Share this project:

Updates