airiskguard 0.1.0

AI Risk Governance Framework — model registry, audit logs, risk dashboards, anomaly detection, regulatory reports, and human review workflows.

airiskguard

AI Risk Governance Framework — model registry, audit logs, risk dashboards, anomaly detection, regulatory reports, and human review workflows for AI applications.

Installation

pip install airiskguard

With optional extras:

pip install airiskguard[fastapi]    # FastAPI integration
pip install airiskguard[flask]      # Flask integration
pip install airiskguard[transformers]  # ML-based hallucination/security detection
pip install airiskguard[dev]        # Development tools

Quick Start

Decorator Pattern

from airiskguard import risk_guard

@risk_guard(checks=["fraud", "compliance"], model_id="payment-v2")
async def process_payment(tx: dict) -> dict:
    return await engine.process(tx)

Middleware Pattern

from fastapi import FastAPI
from airiskguard.integrations.fastapi import add_risk_guard

app = FastAPI()
add_risk_guard(app, config="airiskguard.yaml")

Direct API

from airiskguard import RiskGuard

guard = RiskGuard()
await guard.initialize()

report = await guard.evaluate(
    input_data={"amount": 500, "user_id": "u123"},
    output_data={"approved": True},
    model_id="my-model",
)
print(report.overall_risk, report.overall_score)

Features

• Model Registry — register, version, and manage model lifecycles
• Audit Log — immutable SHA-256 hash-chain audit trail
• Risk Dashboard — aggregate metrics, trends, and JSON export
• Anomaly Detection — IsolationForest + KS drift detection
• Regulatory Reports — GDPR, SOX, EU AI Act reports in JSON + HTML
• Human Review — threshold-based flagging, approve/reject/escalate workflows
• Risk Checkers — fraud, hallucination, compliance, bias, security
• Framework Integration — FastAPI, Flask, ASGI, WSGI middleware

Risk Checkers

Checker	Detects
fraud	Amount anomaly (z-score), velocity abuse, suspicious patterns
hallucination	Fabricated URLs, unverifiable citations, contradictions, NLI
compliance	PII (SSN, email, CC, phone), prohibited content, custom rules
bias	Disparate impact (4/5ths rule), demographic parity, equalized odds
security	Prompt injection (~30 patterns), jailbreak (~20 patterns), encoding attacks

License

MIT