All architectures
Proposed DesignBanking AIOpenStack / hybridGoogle CloudMLOps and CI/CD

Hybrid Banking AI / MLOps — OpenStack Production + GitLab CI/CD + Google Vertex AI

A proposed target design. Components shown are planned or conceptual unless stated otherwise.

The same hybrid pattern on Google Cloud: production inference stays on OpenStack while Vertex AI Pipelines, a model registry and a GKE staging replica support development and evaluation.

Overview

  • Production on OpenStack Kubernetes: FastAPI, a LangGraph credit-appraisal agent with specialist agents, MCP tool gateway, private Llama on vLLM, human approval by a credit officer and Temenos write-back.
  • Data preparation on OpenStack, with de-identification validation (e.g. Presidio / ARX) before encrypted transfer of approved data to Cloud Storage.
  • Vertex AI Pipelines for fine-tuning (Llama / LoRA), evaluation and model registry; staging on GKE / Cloud Run with vLLM, PostgreSQL + pgvector and Neo4j.
  • GitLab CI/CD with quality gates and manual approval; Argo CD deploys approved releases back to OpenStack.
  • A nine-step improvement cycle from production usage to monitoring and feedback.

Business use case

An alternative to the AWS variant for organisations standardised on Google Cloud, with the same rule: only approved, de-identified datasets leave the private cloud.

My role

Architecture design and documentation.