📊 Live System Metrics (last 30 days)
🌟 Our AI Principles
🔒
Privacy by Design
PII masking runs on every query before it reaches any LLM. Named entities, emails, and financial identifiers are replaced with typed placeholders. Redaction events are logged but values are never stored.
🧪
Human in the Loop
All write-back actions (Jira, Slack, SharePoint) require admin approval before execution. Critic scores below 0.40 are blocked automatically; scores 0.40–0.79 trigger human review.
⚖️
Fairness Monitoring
We continuously measure quality equity across departments. A fairness score of 100 indicates identical quality; deviations trigger internal review to identify and correct retrieval disparities.
🔍
Explainability
Every AI answer is traceable: users with appropriate permissions can inspect which document chunks were retrieved, their relevance scores, and how reranking changed the order.
🛡️
Policy Enforcement
Content policies are enforced at both input and output. Queries that trigger policy violations are blocked or flagged before any response is generated or stored.
📋
Audit Logging
All security events, policy violations, and administrative actions are logged immutably to an audit table. Logs are retained and available for compliance review.
🔧 Active Safety Controls
-
🎭
PII Masking (Presidio + regex fallback)Runs on every query and response. Detects names, emails, SSNs, phone numbers, and card numbers.
-
☣️
Toxicity DetectionKeyword-based classifier active by default. Azure Content Safety API available for higher-accuracy enterprise deployments.
-
🤖
LLM-as-Critic ScoringEvery agent response is scored on 5 quality dimensions: relevance, faithfulness, completeness, clarity, and safety. Low scores trigger review or block.
-
💰
Cost GovernancePer-user and per-department budget limits enforced via Redis accumulators. Hard limits block queries; soft limits send alerts.
-
🔐
ABAC Document AccessEvery search result is post-filtered by the user's department, clearance level, and group memberships. Users never see documents outside their access scope.
-
📊
Continuous EvaluationNightly golden dataset evaluation runs automatically. Regression alerts fire to Slack if quality drops more than 5% from the previous run.
—