Skip to content

Feature: Privacy & Governance Layer

The Privacy & Governance Layer is a comprehensive framework that enables safe AI adoption in highly regulated environments (such as government and healthcare) by enforcing strict PII scrubbing and cost controls.

The Innovation: The Privacy Proxy

Instead of sending raw customer data directly to cloud AI providers, Ticket Masala acts as a Secure Gateway:

  • Pre-Redaction: The system detects and scrubs PII before the data leaves the organizational perimeter.
  • Budget Enforcment: Hard and soft caps are applied at the proxy level, ensuring that no single project or user can cause an unexpected bill.
  • Scrubbed Auditing: We maintain a complete history of AI interactions, but only in their redacted form, ensuring GDPR compliance while providing full transparency.

Business Value

The Problem: Compliance Blockers

Many organizations cannot use LLMs because of the risk of PII leakage (GDPR Article 5) and the unpredictable cost of token-based pricing.

The Solution: "Safe-to-Operate" AI

We eliminate the compliance risk by localizing the "Security Perimeter" and providing predictable maintenance costs for the AI subsystem.

Technical Architecture

graph TD
    subgraph "Privacy Proxy"
        Input[Raw Text/PDF] -->|1. Scrub| Scrubber[Local PiiScrubber]
        Scrubber -->|2. Check| Gov[Governance Service]
        Gov -->|3. Approved?| Decision{Budget OK?}
    end

    Decision -->|No| Reject[429 Too Many Requests]
    Decision -->|Yes| LLM[External LLM Service]

    LLM -->|Response| Audit[Scrubbed Audit Log]

Detailed Capabilities

1. PII Scrubber (Regex & Local ML)

The scubber targets patterns critical for compliance without relying on cloud services: - National IDs: (e.g., NISS, VAT numbers). - Contact Info: Emails and Phone numbers. - Financial Info: IBANs and credit card patterns. 

// Example patterns monitored by the Scrubber
private static readonly Regex NissRegex = new(@"\d{2}\.\d{2}\.\d{2}-\d{3}\.\d{2}");
private static readonly Regex EmailRegex = new(@"[A-Z0-9._%+-]+@[A-Z0-9.-]+\.[A-Z]{2,}");

2. Governance Tiers

  • Tier 1 (Individual): Prevents a single agent from dominating the AI resources.
  • Tier 2 (Domain/Tenant): Protects the department budget.
  • Tier 3 (Organization): Global hard stop to prevent critical budget overruns.

3. Compliance Dashboard

Provides real-time visibility into cost projections and "Risk Scores" (the percentage of intercepted PII). This allows DPOs (Data Protection Officers) to verify that the system is working as intended.

Operational Scenarios

Graceful Degradation

If the monthly budget for LLM calls is reached at 95%, the system automatically falls back to "Local Discovery" mode. In this state, LLM suggestions are disabled, but the Enrichment Pipeline continues to show similar historical cases from the local Knowledge Base (RAG-only mode).

Auditing a "Redacted" Discovery

If an agent reports a weird AI suggestion, the admin can look in the AiUsageLogs. They will see the scrubbed input (e.g., "Citizen [NISS_REDACTED] requested a refund"), allowing them to debug the prompt logic without ever seeing the actual citizen's identity.

Success Criteria

  1. Safety: Pass a third-party GDPR audit with zero data leakage findings.
  2. Cost Accuracy: AI spend remains within ± 2% of the configured budget.
  3. Speed: Scrubbing adds less than 50ms latency to the enrichment process.

References