Skip to content

Home

Jump to bottom
AYUSHMAN Mukherjee edited this page Apr 13, 2026 · 1 revision

FairOps Wiki

Real-Time ML Bias Monitoring & Automated Mitigation Pipeline
Google Solution Challenge 2026 · Track: Unbiased AI Decision
Team: Toro Bees

Navigation

Page Description
Home This page — project overview and quick links
Architecture Full system architecture and GCP data flow
SDK Integration How to integrate fairops-sdk in 3 lines
Fairness Metrics All 12 metrics — formulas, thresholds, breach logic
Severity Classification CRITICAL / HIGH / MEDIUM / LOW decision tree
Mitigation Pipeline 10-step Vertex AI KFP v2 DAG explained
Explainability SHAP + Gemini Pro bias narrative generation
Infrastructure Terraform modules, GCP services, IAM setup
Security & Compliance EU AI Act, EEOC, GDPR, DPDPA coverage
API Reference All REST endpoints with request/response schemas
Data Schemas Pydantic v2 schemas — PredictionEvent, BiasAuditResult, MitigationRecord
BigQuery Schema Full DDL for all 5 tables
Spanner Audit Ledger Immutable audit trail design and event types
Demographic Enrichment BISG proxy mode, PII redaction, Cloud DLP
Running Tests Unit, integration, and load test instructions
Sprint Log What was built in each sprint

What is FairOps?

FairOps is a production-grade, GCP-native MLOps platform that continuously monitors deployed ML models for algorithmic bias, automatically triggers a 10-step Vertex AI mitigation pipeline when bias is detected, and generates Gemini Pro-powered plain-English explanations for compliance officers.

The core loop: Prediction events → Cloud Pub/Sub → Dataflow → BigQuery → 12 fairness metrics every 15 min → CRITICAL/HIGH detected → Vertex AI Pipeline auto-triggered → AIF360 mitigation → Model retrained + promoted → Cloud Spanner audit record → Gemini Pro explanation → Looker Studio dashboard

Target domains: Hiring · Credit & Lending · Healthcare Triage · Criminal Justice · Content Recommendation

Quickstart

pip install fairops-sdk
from fairops_sdk import FairOpsClient

client = FairOpsClient(
    project_id="fairops-prod",
    model_id="hiring-classifier",
    model_version="v2.1",
    use_case="hiring",
    tenant_id="acme-corp",
)

client.log_prediction(
    features={"age": 35, "sex": "Male", "education": "Bachelors"},
    prediction={"label": "approved", "score": 0.87, "threshold": 0.5},
)

FairOps monitors your model automatically from this point. No further integration required.

Key Numbers

Metric Value
Fairness metrics computed per audit 12
Audit cycle frequency Every 15 minutes
Bias detection latency < 5 minutes
Mitigation turnaround < 4 hours
Accuracy loss limit (enforced) < 2%
Regulatory frameworks covered EU AI Act · EEOC · GDPR Art.22 · India DPDPA

Links

Built by Toro Bees · Google Solution Challenge 2026

Clone this wiki locally