The problem TARK solves

The Problem TARK Solves

Cyber scams are growing rapidly in both scale and sophistication.
Users are constantly targeted through phishing links, fake KYC updates, OTP fraud, and impersonation messages.

The biggest challenge is not just identifying scams, but understanding them and taking action in time.

❗ Key Problems

1. Lack of Awareness

Most users cannot distinguish between:

• Real vs fake bank messages
• Legitimate vs phishing links
• Safe vs malicious requests

Scammers exploit this gap using urgency, fear, and authority.

---

2. Generic & Unreliable Tools

Existing solutions:

• Provide generic warnings
• Rely heavily on black-box
• Do not explain why something is a scam

👉 This reduces trust and usability

---

3. No Actionable Output

Even if a scam is detected:

• Users don’t know what to do next
• Reporting to authorities is complicated
• No structured evidence is generated

---

4. Privacy Risks

Scam messages often contain:

• Phone numbers
• Aadhaar details
• Banking information

Sending this data to AI systems can lead to privacy violations

💡 How TARK Solves This

🧠 1. Intelligent Scam Analysis

TARK analyzes messages using a multi-layer AI system to:

• Detect scam type (KYC, OTP, phishing)
• Identify manipulation techniques (urgency, fear, authority)
• Provide confidence and reasoning

---

🔍 2. Explainable Output

Instead of just saying “this is a scam,” TARK explains:

• Why it is a scam, exposes links
• What techniques are being used
• What risks are involved

👉 Builds user trust and awareness

---

🔗 3. Link Safety Detection

• Extracts URLs from messages
• Flags suspicious domains
• Identifies phishing patterns

---

📄 4. One-Click Report Generation

• Generates a professional, structured PDF report
• Includes analysis, evidence, and blockchain proof
• Allows users to directly report to cybercrime authorities via Gmail

---

🔒 5. Privacy-First Processing (🤏 Smolify)

• Masks sensitive data before analysis
• Ensures no personal information is exposed
• Enables safe AI usage

---

🔗 6. Tamper-Proof Evidence (Algorand)

• Stores hash of scam data on blockchain
• Provides verifiable proof for investigations
• Ensures integrity of reports

---

🚀 Real-World Use Cases

• 🧑‍💻 Users verifying suspicious messages before acting
• 🏛️ Citizens reporting scams to cybercrime departments
• 🏦 Banks detecting phishing patterns
• 📱 Integration into apps for real-time scam detection
• 🎓 Awareness and training for digital safety

---

🌟 Impact

TARK transforms the process from:

❌ Confusion → ❌ Fear → ❌ Inaction

to:

✅ Detection → ✅ Understanding → ✅ Action

---

TARK doesn’t just detect scams — it empowers users to respond intelligently and safely.

Challenges we ran into

From Challenges We Ran Into to Finding their Solutions

---

🔹 1. RAG Not Returning Meaningful Violations

Problem: Retrieved chunks lacked source/context, leading to vague outputs like “Violation: NONE”
Solution:

• Added structured metadata (title, category, violation) to Pinecone
• Implemented hybrid fallback (hardcoded policy) when retrieval fails

---

🔹 2. Gmail Link Breaking (Report Feature)

Problem: Email links failed due to extremely long, unencoded body content
Solution:

• Proper URL encoding using urllib.parse.quote
• Reduced email body to summary only, moved full data to PDF

---

🔹 3. API Rate Limits (Embeddings)**

Problem: Batch embedding requests caused rate limit errors
Solution:

• Implemented batching + retry logic
• Added delays sleep between requests

---

🔹 4. Privacy Concerns in User Input

Problem: Scam inputs contained sensitive data (phone, Aadhaar, etc.)
Solution:

• Integrated 🤏 Smolify as a local masking layer before processing

---

🔹 5. “LLM Wrapper” Perception

Problem: Risk of being seen as just an LLM-based system
Solution:

• Introduced dataset-driven vector search (RAG)
• Positioned LLM only for explanation, not detection

---

🔹 6. Inconsistent Scam Classification

Problem: Pure LLM outputs were sometimes inconsistent
Solution:

• Added retrieval-based grounding + heuristic checks
• Ensured deterministic fallback logic

---

🔹 7. Lack of Trust in Generated Reports

Problem: Reports could be questioned for authenticity
Solution:

• Integrated Algorand blockchain to store hash + TX ID
• Made reports tamper-proof and verifiable

---

✅ Key Learning

Building reliable AI systems requires combining:
ML + structured data + fallback logic + trust layers, not just LLMs.