06 MLOPS Regulated sectors Critical

Data Engineering / MLOps — ML Pipeline in Production

ML pipeline with fraud detection model, a high-risk system under the EU AI Act.

◻ intake ★★★

△ architect ★★★

◇ vigil ★★☆

⬡ licit ★★★

Architecture diagram

Context

ML pipeline for fraud detection with Feature Store (Feast), Training Pipeline (Airflow), Model Registry (MLflow), and Serving (FastAPI). The EU AI Act directly applies to AI models in production.

Flow with 4 tools

◻ Phase 01 — intake

From notebooks to specs

Converts research notebooks and MLOps RFCs into executable specifications.

◻ intake

intake init "Fraud Detection Model to Production" \
  --source notebooks/fraud_detection_v3.ipynb \
  --source docs/mlops-infrastructure-rfc.md \
  --source gdocs://ml-team/fraud-model-kpis \
  --mode enterprise

△ Phase 02 — architect

Implement MLOps stack

Automated pipeline to bring the model to production.

△ architect

architect pipeline pipelines/ml-to-production.yaml \
  --var model="fraud_detection"

⬡ Phase 03 — licit

EU AI Act for ML

Mandatory FRIA for fraud detection model. Annex IV with model architecture, training data, and metrics.

⬡ licit

licit init && licit trace
licit fria
licit annex-iv
licit report --format html
licit verify --min-score 80

Why licit is critical here

A fraud detection model is a high-risk AI system under the EU AI Act. licit evaluates the 11 relevant articles, generates the FRIA documenting the impact on financial rights, and produces the mandatory Annex IV.