LUMINA

The call begins. The pressure rises. The trap tightens. LUMINA detects digital-arrest coercion in real time—and when the victim can't call for help, LUMINA does.

Comment

LUMINA — Digital Incident Protection

💡 Inspiration

Your phone rings. You answer.

The caller sounds professional. They know your name.

Then:

“Your identity has been linked to a criminal case.”

“Do not disconnect.” “Do not tell your family.” “You have to act immediately.”

At that moment, the attack is no longer targeting only a device.

It is targeting a person's decisions.

The victim could be anyone — you, your parents, grandparents, friends, siblings, or colleagues.

India recorded more than 6.58 million financial-fraud complaints and over ₹55,050 crore in reported losses between 2021 and 2025.

In 2026, the CBI reported a case in which a senior citizen was coerced into transferring ₹25.65 lakh after being threatened with a digital arrest.

These incidents made me ask:

What happens during the minutes when manipulation is still unfolding?

Existing cybersecurity protects different parts of the problem — devices, accounts, transactions, caller identity, suspicious activity, and post-incident reporting.

But the human decision being manipulated during the conversation remains a critical protection challenge.

That question became LUMINA.

🛡️ What it does

LUMINA is a digital incident protection system combining AI/ML and cybersecurity to help a person recognize, interrupt, and respond to coercive digital conversations.

The focus is not simply:

“Is this caller a scammer?”

It is:

“What is happening to this person, and what safer action can happen next?”

Through an authorized audio pathway, LUMINA converts speech into text and analyzes combinations of:

AUTHORITY → THREAT → URGENCY → ISOLATION → SENSITIVE REQUEST → FINANCIAL PRESSURE

A single word does not determine risk. LUMINA is designed to understand the pattern and progression of manipulation.

Core flow

AUTHORIZED AUDIO

↓

SPEECH-TO-TEXT

↓

AI/ML CONVERSATION UNDERSTANDING

↓

STRUCTURED EVIDENCE

↓

CYBERSECURITY SAFETY ENGINE

↓

INTERVENTION

↓

TRUSTED HUMAN ASSISTANCE

↓

CONTAIN → RECOVER

I am developing the conversational AI/ML layer around a DeBERTa-v3-base multi-label classifier, designed to identify multiple coercion tactics within the same conversation.

AI interprets. The safety layer decides.

⚠️ During the attack

When pressure is identified, LUMINA can provide quiet interventions such as:

  • “Take a moment before continuing.”
  • “Pause before sharing sensitive information.”
  • “Consider verifying through an independent channel.”

🆘 I'M TRAPPED — GET HELP

If the person recognizes the danger first, they can immediately request help through I'M TRAPPED — GET HELP.

LUMINA activates the configured trusted-contact workflow, giving the person a direct path to human support.

The trusted contact receives relevant incident context, such as:

Possible coercive digital incident Pressure indicators detected User requested assistance Recommended action: Contact the user and help them verify independently.

Sensitive secrets such as passwords, OTPs, PINs, and authentication codes are not included.

A structured Help Story organizes the incident into:

FACT → USER REPORT → INFERENCE → UNKNOWN → ACTION

👥 Target Users

LUMINA is designed for people who may face coercive digital conversations, including:

  • Individuals receiving suspicious or high-pressure calls
  • Families and trusted contacts who can assist during an incident
  • People with different levels of digital literacy who may need clearer, calmer guidance

The goal is to support anyone facing digital manipulation, rather than a single age group.

🔴 Why the “DURING” stage matters

Protection is distributed across different stages:

BEFORE → prevention and detection DURING → active manipulation AFTER → reporting, containment, recovery

LUMINA focuses on the DURING stage — while the attacker is actively trying to influence the person's next decision.

🧩 How I built it

I am building LUMINA combining AI/ML, cybersecurity, and human-centered safety design.

The system is partially built and actively under development, using:

  • Android / Kotlin — mobile safety experience
  • Python / FastAPI / Pydantic — backend and safety services
  • faster-whisper — speech-to-text
  • XGBoost — risk-scoring layer
  • DeBERTa-v3-base — conversational classification architecture
  • React / TypeScript / Vite — web interface
  • SQLite — current persistence
  • AES-256-GCM / HMAC-SHA256 / Android Keystore — security and protected data handling
  • Vercel / Render — deployment infrastructure

The architecture follows:

CONVERSATION → AI/ML → EVIDENCE → SAFETY POLICY → INTERVENTION

The production conversational model still requires properly consented, privacy-cleared, expert-annotated real-world data and rigorous validation before production deployment.

LUMINA uses authorized audio access and user permissions, rather than hidden call interception.

⚙️ Challenges I ran into

The central challenge was:

How do I intervene without creating unnecessary panic?

Legitimate conversations can contain words such as “police,” “payment,” or “urgent.”

So LUMINA is designed to consider multiple indicators, context, and escalation rather than individual keywords.

I also had to balance AI intelligence with user control, leading to:

UNDERSTAND → EVALUATE → INTERVENE

🏆 Accomplishments I am proud of

I am proud that LUMINA connects:

AI UNDERSTANDING → INTERVENTION → TRUSTED HUMAN → CONTAINMENT → RECOVERY

I am especially proud of the I'M TRAPPED — GET HELP pathway because the person can request human assistance without waiting for AI detection.

They can ask for help themselves.

📚 What I learned

Building LUMINA changed how I think about both AI and cybersecurity.

A cyberattack does not always begin with malware.

Sometimes it begins with trust.

An attacker may not need to break into a device if they can persuade the person holding it to make the harmful decision themselves.

I learned that AI can help cybersecurity move beyond detecting technical threats and toward understanding human-facing attacks.

A trusted human can be part of a security system.

🌐 What's next for LUMINA

Multilingual

Expand conversational understanding across multiple languages and code-switched speech.

Location assistance

With explicit user consent and permissions, allow trusted contacts to receive the user's location during an active help request.

Stronger understanding

Improve pressure, escalation, speaker awareness, and real-world validation.

Connect → Contain → Recover

Expand trusted-contact assistance and strengthen incident recovery.

🎯 The Final Thought

A scammer may need only a few minutes of your trust.

LUMINA is designed for those minutes.

When the attacker says:

“Don't tell anyone.”

LUMINA gives the person another choice:

“Get someone you trust.”

The attack may be digital.

The consequence is human.

Built With

Share this project:

Updates

Submission history