Documentation
Everything you need to integrate AI compliance monitoring into your applications.
Getting Started
Create an account
Sign up for a free Neural Detective account to access the monitoring dashboard and API keys.
Generate an API key
From the monitoring dashboard, create an API key. Include it as a Bearer token in all API requests.
Make your first API call
Record a decision, register a model, or scan text for safety. See the example below.
curl -X POST https://api.neuraldetective.com/api/v1/decisions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model_id": "your-model-uuid",
"inputs": { "applicant_id": "eu-7291" },
"output": { "decision": "approved", "confidence": 0.87 }
}'Why It Matters
AI compliance isn't optional — and the cost of getting it wrong is rising fast.
EU AI Act Compliance
High-risk AI systems must demonstrate continuous monitoring and bias testing. Fines reach up to 35 million euros or 7% of global turnover.
Silent Model Degradation
Drift degrades predictions silently. A lending model can approve bad loans for months before anyone notices.
Bias & Discrimination Risk
Disparate impact liability doesn't require intent. The 4/5ths rule is used by regulators on both sides of the Atlantic.
Content Safety & Trust
PII leakage, toxic output, or missing disclaimers in AI-generated content create brand and legal risk.
What We Measure
Three capabilities that cover the full AI risk surface — from data drift to content safety.
Drift Detection
Catch distribution shifts before they degrade predictions.
Population Stability Index
Compares live input distributions to training baselines. A value above 0.25 indicates significant shift.
Jensen-Shannon Divergence
Symmetric divergence measure on a 0–1 scale. Works well for categorical features.
Wasserstein Distance
Measures the magnitude of distribution shift for continuous features.
Confidence-Based Performance Estimation
Estimates accuracy without ground-truth labels using model confidence scores.
Bias Monitoring
Quantify fairness and produce audit-ready evidence.
Disparate Impact Ratio
Compares favourable outcome rates across groups. The 4/5ths rule threshold is a ratio of 0.8 or above.
Group Outcome Rates
Per-attribute breakdown (gender, age group, etc.) for audit evidence and regulatory reporting.
Safety Scanning
A three-tier pipeline that balances speed with depth.
Tier 1 — Rules & PII
Regex patterns and PII detection. Runs in approximately 5 ms.
Tier 2 — ML Classifiers
Toxicity and content-policy classifiers. Runs in 50–200 ms.
Tier 3 — LLM Judge
Optional large-language-model evaluation for nuanced cases. Runs in 500 ms – 3 s.
What it detects
Toxic language, hate speech, PII (emails, phone numbers, SSNs), prompt injection attempts, missing safety disclaimers, and off-topic content.
How You Stay Informed
From dashboards to webhook events — know the moment something changes.
Real-Time Dashboard
Drift trends, alert history, and scan verdicts — filterable by model and time range.
Compliance Reports
Five-section, PDF-exportable reports covering executive summary, model health, safety, alerts, and usage.
Configurable Alerts
Set per-metric warning and critical thresholds with cooldown periods to avoid alert fatigue.
Webhook Notifications
HMAC-signed alert.created events delivered to your endpoint for Slack, PagerDuty, or custom integrations.
API Overview
The Neural Detective API is organized around three capabilities.
Decisions & Models
Register AI models, record decisions, and upload training baselines for drift detection.
Enables drift detection and audit logging required by EU AI Act Article 12.
/api/v1/decisions
Record an AI decision for monitoring and drift detection.
/api/v1/models
Register a new model for monitoring.
/api/v1/models
List all monitored models.
/api/v1/models/:id
Retrieve details for a single model.
/api/v1/models/:id/baselines
Upload training baseline for drift detection.
Safety Scanning
Scan AI-generated text for harmful, toxic, or non-compliant content in real-time or asynchronously.
Prevents PII leakage and toxic content from reaching your users in production.
/api/v1/scans
Scan text for harmful content (sync).
/api/v1/scans/async
Scan text for harmful content (async).
Bias Monitoring
Attach protected attributes to decisions to enable disparate impact analysis and fairness reporting.
Documents fairness for the 4/5ths rule and EU AI Act non-discrimination requirements.
/api/v1/protected_attributes
Attach protected attributes for bias monitoring.
Authentication
All API requests require a Bearer token. Include your API key in the Authorization header of every request.
Authorization: Bearer YOUR_API_KEYWebhooks
Receive real-time notifications when alerts are triggered — such as drift threshold breaches or bias detection events. Configure webhook endpoints from the monitoring dashboard.
alert.created
Fired when a new alert is created (e.g. drift threshold exceeded, bias detected).
Ready to get started?
Create a free account to access the full API reference, generate API keys, and start monitoring your AI models.