Cleir

AI protection platform. Personal mode catches manipulation in conversations. Business mode validates visas, analyzes contracts, detects fraud. 8 agents. Real APIs. Catches what you miss.

Comment

Inspiration

What it does## Inspiration

My cousin lost £400 on a visa rejection. One missing document. Her passport had 5 months validity the UK requires 6. Nobody told her.

My uncle almost wired £2,000 to someone impersonating his bank.

My friend signed an employment contract with a 2-year non-compete she didn't understand. She couldn't work in her industry for 2 years after leaving.

Every one of these was preventable. The information existed. The red flags were there. But nobody was watching.

I built Guardian because the people I care about keep getting hurt by things AI should catch.

What it does

Guardian is an AI protection platform. Two modes. One mission: catch what you miss.

Personal Mode — Relationship Protection:

  • Upload screenshots of conversations
  • Detects manipulation tactics (gaslighting, guilt-tripping, love bombing, etc.)
  • Translates what they said vs what they meant
  • Provides boundary scripts exactly what to say back
  • Tracks patterns over time across conversations
  • Auto-detects: relationship manipulation, scams, or self-analysis needed

Business Mode — Career & Travel Protection:

🛂 VisaLens — Document Validator

  • Upload passport, bank statements, employment letters
  • Validates against real visa requirements (Diaspora AI API)
  • Shows approval likelihood percentage
  • Flags missing documents before you apply

📜 LegalLens — Contract Analyzer

  • Upload any contract (employment, NDA, lease)
  • Detects red flags (non-competes, IP assignment, arbitration clauses)
  • Translates legal jargon to plain English
  • Generates "must negotiate" list

🎣 B2B ScamShield — Fraud Detection

  • Upload suspicious emails or invoices
  • Detects CEO fraud, payment redirects, domain spoofing
  • Verifies sender legitimacy
  • Provides action checklist

✈️ TripGuard — Multi-City Travel Planner

  • Plan multi-destination trips
  • Checks visa requirements for every stop
  • Generates combined document checklist
  • AI-powered itinerary generation with Google Places data
  • Interactive maps with real ratings, photos, hours

How I built it

Architecture:

User Input → Auto-Detection Router → Specialized Agent Pipeline → Synthesis → Response

Personal Mode: 5-Agent Pipeline
├── Extractor (vision + text parsing)
├── Classifier (35+ manipulation tactics)
├── Psychologist (relationship health scoring)
├── Defender (boundary script generation)
└── Guardian (synthesis + voice output)

Business Mode: 4 Specialized Agents
├── VisaLens (document validation + DA API)
├── LegalLens (contract analysis + red flags)
├── ScamShield (fraud pattern detection)
└── TripGuard (multi-city planning + Places API)

Technical Stack:

  • Frontend: Next.js 14, React, TypeScript, TailwindCSS, Framer Motion
  • Backend: Next.js API Routes, Server Actions
  • AI: Gemini 3 Flash (multimodal reasoning, auto-detection, analysis)
  • Voice: ElevenLabs TTS with animated avatar
  • Auth: Supabase Auth (email/password + GitHub OAuth)
  • Database: Supabase PostgreSQL (analysis history, pattern tracking)
  • External APIs:
    • Diaspora AI Visa API (real visa requirements)
    • Google Places API (location enrichment)
    • Google Maps API (interactive itineraries)
  • Deployment: Vercel

The Auto-Detection System:

Users don't select which agent to use. They just upload. Gemini analyzes the content and routes to the right pipeline:

  • Passport detected → VisaLens
  • Contract detected → LegalLens
  • Suspicious email → ScamShield
  • Travel itinerary → TripGuard
  • Conversation screenshot → Personal Mode (relationship/scam/self-analysis)

One interface. AI figures out the rest.

Diaspora AI Integration:

Guardian connects to Diaspora AI a travel booking platform I built for the African diaspora. The Visa API provides real requirements:

  • Visa types by nationality/destination
  • Document requirements
  • Financial thresholds
  • Processing times
  • Direct booking links

This isn't hallucinated data. It's production API calls to a real platform serving real users.

Challenges I ran into

1. Manipulation Taxonomy How do you classify emotional manipulation? I researched psychology literature and built taxonomies:

  • 15 relationship manipulation tactics
  • 12 scam patterns
  • 8 self-analysis patterns

Each with severity levels, examples, and response strategies.

2. False Positives in Relationship Analysis Early versions flagged normal disagreements as manipulation. Solution: Added context weighting — a single "you always" isn't gaslighting, but combined with other patterns it might be.

3. Contract Clause Extraction Contracts have inconsistent formatting. Some use numbered sections, some don't. I trained the prompt to handle multiple formats and extract clauses regardless of structure.

4. Streaming Locations for TripGuard First version asked Gemini for streaming/travel data. It hallucinated. Solution: Only use AI for reasoning, never for data retrieval. All factual data comes from APIs.

5. Voice Synthesis Latency ElevenLabs adds ~2 seconds. For personal mode (emotional content), users want immediate feedback. Solution: Stream text response first, voice follows. Users read while audio loads.

6. Mobile Sidebar UX Desktop sidebar doesn't work on mobile. Built a full overlay system with swipe gestures (swipe from left edge to open, swipe left to close).

Accomplishments that I'm proud of

  • 8 specialized AI agents working in orchestration
  • 35+ detection patterns across manipulation, scams, and contracts
  • Real API integrations not mocked data
  • Auto-detection routing users never pick an agent
  • Production-ready error handling handles edge cases gracefully
  • Built in 72 hours while working full-time cleaning job (overnight shifts)
  • Connects to my own startup Diaspora AI integration shows real-world applicability

What I learned

  • Multi-agent orchestration is more powerful than single-prompt systems
  • Auto-detection removes friction don't make users think
  • The "Action Era" means AI should DO things, not just analyze
  • Real APIs > hallucinated data (always)
  • Building for your community hits different every feature came from a real story

What's next for Guardian

  • Evidence Vault: Save and organize manipulation evidence for legal/therapy use
  • Lawyer Matching: Connect users with lawyers for flagged contracts
  • Real-time Email Monitoring: Gmail/Outlook integration for automatic scam detection
  • Visa Application Tracker: Track application status across multiple countries
  • Group Analysis: Analyze group chats for toxic dynamics
  • Therapist Integration: Share pattern reports with mental health professionals
  • Browser Extension: Analyze any conversation on any platform ```

Built With

Share this project:

Updates