VerifyLive

💡 Inspiration

In an era where GenAI can create hyper-realistic Deepfakes in seconds, traditional identity verification methods (simple face matching) are obsolete. We witnessed a 300% rise in identity fraud targeting fintechs and hiring platforms.

However, the tipping point for us was the introduction of Lei Felca (Law 14.811/2024) in Brazil, which mandates strict digital protection for children. We realized that current tools aren't just failing to detect fakes; they are failing to protect vulnerable groups. We wanted to build a system that is Forensic by Design.

🛡️ What it does

VerifyLive is a next-gen biometric security ecosystem that goes beyond simple face matching.

1. Active Liveness: Users perform a sequence of 5 random physical challenges (e.g., "Turn right", "Zoom in", "Show ID").
2. Gemini 3 Forensics: We don't just compare pixels. We send the session frames to Gemini 3 Multimodal, which acts as a forensic expert, analyzing texture, micro-expressions, and lighting inconsistencies to detect Deepfakes.
3. Compliance Guard: The system automatically estimates age and enforces Lei Felca (blocking under-16s without parental tokens) and LGPD (auto-deleting sensitive biometric data after 24h).

⚙️ How we built it

We architected the solution using a Security-First approach:

• Frontend: Built with Next.js 15 and MediaPipe for client-side face landmark detection (zero-latency feedback).
• The Brain (AI): We leveraged Google Vertex AI to access Gemini 3 Pro. We crafted complex system prompts that instruct the model to look for "artifacts of generation" rather than just facial similarity.
• Orchestration: Used Context7 MCP to manage the data flow context between the user session and the AI analysis.
• Backend & Storage: Supabase (PostgreSQL) for relational data and Google Cloud Storage for ephemeral evidence storage.
• Auditing: Integrated Google Cloud Logging to create an immutable audit trail for every verification attempt, crucial for legal validity.

🧠 Challenges we ran into

• Latency vs. Accuracy: Balancing the deep analysis of Gemini 3 with the user need for speed. We solved this by using MediaPipe for immediate UI feedback while Gemini processes the forensic analysis asynchronously via Edge Functions.
• Prompt Engineering for Forensics: Teaching an LLM to "think like a forensic expert" required multiple iterations. We had to move from generic prompts to specific instructions about "skin texture analysis" and "environmental lighting consistency".
• Strict Compliance: Implementing "Privacy by Design" meant we couldn't just save everything. We had to build strict TTL (Time-To-Live) policies to ensure data is wiped automatically.

🏆 Accomplishments that we're proud of

• Successfully integrating Gemini 3 not as a chatbot, but as a visual reasoning engine for security.
• Building a fully functional "Anti-Deepfake" pipeline that catches injection attacks.
• Achieving a 100% Compliance Score with Brazilian laws (LGPD/Lei Felca) within the hackathon timeframe.

📚 What we learned

We learned that Gemini 3 has incredible capability for visual nuance. It can spot subtle inconsistencies in a "fake" video feed (like a background that doesn't move correctly with the head) that traditional CV algorithms miss. We also learned the power of MCP (Model Context Protocol) in keeping our development context clean and organized.

🚀 What's next for VerifyLive

• SaaS White-label: Launching as an API for EdTechs and Fintechs.
• Edge AI: Moving lighter forensic models to run directly on the device using Gemini Nano.
• Global Compliance: Expanding the rule engine to cover GDPR (Europe) and COPPA (USA).