• 

  Dashboards

• 

  Demo chat

AI Governance Platform - Enterprise AI Safety & Compliance

Inspiration

As AI adoption accelerates across enterprises, the need for robust governance and safety measures has become critical. Organizations deploying AI agents face challenges around:

• Data Privacy: Preventing accidental exposure of PII (Personally Identifiable Information)
• Content Safety: Filtering toxic or harmful content in AI interactions
• Security: Detecting prompt injection attacks that could manipulate AI behavior
• Compliance: Maintaining audit trails for regulatory requirements

I built this platform to demonstrate how Google's Gemini 3 can power a comprehensive AI governance solution that runs entirely on Google Cloud.

What It Does

The AI Governance Platform provides real-time guardrails for AI agents:

How I Built It

Tech Stack

• Backend: FastAPI (Python)
• AI Model: Gemini 3 Flash Preview via Google AI Studio
• Database: Google BigQuery for audit logs
• Deployment: Google Cloud Run (serverless)
• Frontend: Vanilla JS with Chart.js for dashboard

Gemini 3 Features Used

1. System Instructions - Consistent AI behavior across all guardrail checks
2. Safety Settings - Native toxicity filtering with configurable thresholds
3. Structured Output - JSON responses for reliable parsing
4. Token Counting - Cost tracking and usage estimation

Key Code Snippet - Toxicity Detection

async def check_toxicity(self, text: str) -> ToxicityResult:
    """Use Gemini's safety features to detect toxic content"""
    prompt = f"""Analyze this text for toxicity:
    Text: {text}

    Return JSON with: is_toxic, toxicity_score (0-1), categories"""

    response = await self.model.generate_content_async(
        prompt,
        safety_settings=self.safety_settings
    )
    return ToxicityResult(**json.loads(response.text)) """

### Challenges I Faced
1. Cloud Run Secret Management
Getting the Gemini API key to work in Cloud Run was tricky. The secret was being injected as an environment variable, but pydantic-settings wasn't picking it up.

Solution: Added fallback to check both pydantic settings AND os.environ.get('GEMINI_API_KEY') directly.

2. API Quota Limits
Hit the free tier limits during development (15 RPM, 1000 RPD).

Solution: Implemented regex-based fallback detection that works even when Gemini is unavailable, ensuring the platform remains functional.

3. Model Availability
Gemini 3 Flash Preview is a cutting-edge model with limited documentation.

Solution: Tested extensively and implemented graceful degradation for unsupported features.

What I Learned:

Gemini 3's Safety Features are powerful - the native toxicity filtering catches nuanced harmful content that regex patterns miss
Serverless architecture with Cloud Run makes deployment and scaling effortless
BigQuery is excellent for audit logging - the streaming inserts handle high throughput while enabling complex compliance queries
Graceful degradation is essential - always have a fallback when external APIs might be unavailable

What's Next:

Multi-model support - Add support for other LLMs (Claude, GPT-4)
Custom policies - Allow enterprises to define their own guardrail rules
Real-time alerts - Slack/email notifications for security events
Fine-tuned detection - Train custom models for domain-specific PII

Try It Out

Live Demo: https://ai-governance-platform-902023244402.us-central1.run.app
API Docs: https://ai-governance-platform-902023244402.us-central1.run.app/docs
Dashboard: https://ai-governance-platform-902023244402.us-central1.run.app/dashboard

Built With

• fastapi
• gcp
• google-bigquery
• google-cloud-run
• python