MLOps architecture for the bank
Credit, fraud, AML, propensity models — dozens in production. MLOps is the operating loop for lifecycle with regulatory traceability.
Discuss Your ChallengeWhy the bank needs MLOps
Banking models have higher stakes than marketing models: credit decisions ($), fraud blocks (UX), AML triggers (regulator). Without MLOps — models age unnoticed, accuracy falls, regulator audit fails.
Structural elements
Feature store. Centralised, versioned. Online (inference) and offline (training) consistency.
Model registry. Versioned models, metadata, status (dev/staging/prod/archived).
Training pipeline. Reproducible — data → preprocessing → train → evaluate → registry.
Deployment. Standardised path: shadow → canary → full with auto-rollback.
Monitoring. Prediction distribution, feature drift, accuracy (ground truth), business KPI.
Governance. Approval workflow, audit trail, explainability, fairness monitoring.
Banking-specific requirements
Regulatory traceability. Every model decision must be reproducible on demand. Versioning critical.
Bias monitoring. Disparate impact analysis mandatory for credit, AML.
Explainability. SHAP / LIME for regulator inquiries.
Model risk management. Banks under Basel/IFRS9 have formal model validation requirements.
Champion/challenger. Active production model and continuously evaluated alternative.
Where it usually breaks
Train-serve skew. Features computed differently in train vs production.
Models in Jupyter — irreproducible.
Deployment through ticket — slow, irregular updates.
No monitoring — accuracy decay months unnoticed.
No registry — production model is “that file”.
Bias not checked — regulator complaint.
Operating model
Owner — Head of Model Risk / Head of Data Science with infra mandate.
Teams: ML platform, data science, ML engineering, model governance.
Routine — weekly production model review.
Related
- /en/architecture/banking-cdp-architecture/ — feature source
- /en/architecture/banking-realtime-decisioning/ — model consumer
- /en/insights/banking-model-risk-management/ — model risk
- /en/insights/banking-ai-governance/ — AI governance
What else is worth exploring
Topics from the same area we usually explore together
CRM
Not an off-the-shelf CRM, but a properly built customer management contour — from first contact to loyalty.
→SolutionBI
Analytics is not pretty charts on the wall. It's the answer to 'why?' before the problem becomes a loss.
→SolutionContact Center
The contact center is not a phone station — it's the point where a client decides: stay with you or leave. The question is how it's built…
→SolutionOnboarding
Onboarding is your company's first impression. If it takes 5 days and 12 paper forms, there won't be a second impression.
→I do not just write about this. I can come in, examine your situation and design a solution for your specific landscape.
Discuss applying this →Ready to discuss your challenge?
Tell me what's not working or what needs to be built. First conversation — no obligations.
Usually respond within a few hours