GuardKall

GuardKall is an AI call firewall that answers unknown callers for you, verifies who they are, and only lets real people reach you.

Comment

Project Story: GuardKall

Inspiration

Spam and scam calls aren’t just annoying — they’re dangerous. Every day, people lose money, privacy, and peace of mind to social engineering attacks where scammers impersonate banks, delivery services, or even government agencies. What makes this worse is that traditional defenses rely heavily on caller ID, number scanning, or name matching, which scammers easily bypass by spoofing numbers or rotating new ones.

What inspired us was a simple realization:

Scams don’t succeed because of phone numbers — they succeed because of human manipulation.

So instead of asking “Who is calling?”, we decided to ask a better question:

“Is this conversation safe right now?”

About the Project

GuardKall is an AI-powered call interception and scam detection system that focuses on behavioral analysis, not just identity verification.

When an unknown or suspicious call is forwarded to GuardKall, the system:

  • Analyzes the live conversation
  • Detects scam patterns such as urgency, threats, financial pressure, or secrecy
  • Intercepts and blocks high-risk calls
  • Provides the user with a clear summary and assessment
  • Learns continuously to improve future detection

To power this at scale, we use Snowflake as our data backbone — not for static number lookups, but to identify scam behavior patterns across calls, even when scammers change phone numbers.

How We Built It

Instead of relying on phone number lookups, our system focuses on how the caller behaves during the conversation.

Core Components

AI Call Analysis Layer

Extracts conversational signals such as urgency, coercion, payment requests, impersonation cues, and escalation patterns.

Scam Probability Scoring Engine

Each call is given a scam probability based on detected behaviors rather than a fixed risk score. The system looks for common scam signals—such as urgency, threats, impersonation, and financial requests—to estimate how likely a call is fraudulent. Using probabilities instead of fixed rules allows the system to adapt to new scam patterns and respond intelligently with warnings, call interception, or blocking based on confidence level.

Snowflake Data Platform

Used to store anonymized call data and scam patterns, enabling us to:

  • Detect repeated scam scripts across different phone numbers
  • Recognize coordinated scam attempts early
  • Continuously improve detection using real-world data

User-Focused Frontend

Designed to be simple and non-technical. Users don’t need to make decisions mid-call — GuardKall acts automatically and explains afterward.

Challenges We Faced

One of our biggest challenges was avoiding false positives while still acting quickly. Blocking legitimate calls can be just as harmful as missing scams, so we had to carefully balance sensitivity and precision.

Another major challenge was data responsibility. Since calls contain sensitive information, we designed our system to:

  • Store only anonymized signals that capture scam behavior
  • Avoid raw personal identifiers
  • Focus on patterns, not people

Finally, building a system that feels protective rather than intrusive pushed us to rethink traditional security UX. The goal was trust, not fear.

What We Learned

  • Identity-based security is fragile when attackers can fake identity
  • Behavioral signals are far harder to disguise than phone numbers
  • Scalable protection requires shared intelligence, not isolated filters
  • Good security UX should reduce user effort, not add to it

Most importantly, we learned that real safety doesn’t come from more labels — it comes from understanding intent in real time.

What’s Next

We plan to expand GuardKall to support:

  • SMS and email scam detection
  • Community-based scam alerts
  • Advanced campaign tracking across regions
  • Accessibility-focused features for elderly and vulnerable users

Our long-term vision is to turn GuardKall into a collective defense layer — where one detected scam helps protect thousands of others.

Built With

Share this project:

Updates