Proposed DesignBanking AIOpenStack / hybridGoogle CloudMLOps and CI/CD
Continuous AI Improvement Workflow — OpenStack to Google Cloud and Back
A proposed target design. Components shown are planned or conceptual unless stated otherwise.
The ten-step improvement loop on Google Cloud: approved, anonymized data goes to Cloud Storage, Vertex AI trains and evaluates, GKE stages, and releases return to OpenStack.
Overview
- On OpenStack: production usage and monitoring, case selection with human feedback, de-identification (e.g. Presidio / ARX) and data-quality approval.
- On Google Cloud: encrypted upload to Cloud Storage with IAM and VPC Service Controls, baseline reproduction, Llama / Gemini LoRA fine-tuning, RAG and agent evaluation in Vertex AI, and staging validation on GKE against business metrics.
- In GitLab: automated tests, security and compliance checks, model evaluation review, human approval and Helm release packages in Artifact Registry.
- Back on OpenStack: Argo CD GitOps sync, canary/regression tests, RAG re-indexing, promote or roll back, and monitoring for the next cycle.
Business use case
Same principle as the AWS variant: live customer data stays private; only approved, de-identified or synthetic data is used for training.
My role
Architecture design and documentation.