All architectures
Proposed DesignBanking AIOpenStack / hybridAWSMLOps and CI/CD

Hybrid Banking AI / MLOps — OpenStack Production + GitLab CI/CD + AWS SageMaker

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

Credit-appraisal AI runs on a private OpenStack Kubernetes cluster; GitLab is the MLOps control plane; AWS SageMaker handles training, evaluation and a staging replica. Only approved, de-identified data moves to AWS.

Overview

  • Production (OpenStack / Kubernetes): FastAPI backend, a LangGraph credit-appraisal agent coordinating credit-scoring, asset-appraisal, legal and fraud agents, an MCP tool gateway, private Llama inference on vLLM, credit rules and approval by a credit officer, and decision write-back to core banking through Temenos APIs. Data layer: PostgreSQL + pgvector, Neo4j and Ceph object storage; Argo CD syncs approved releases.
  • GitLab CI/CD: build, unit, security/compliance and integration tests, release quality gates, manual release approval, GitOps manifests and a private artifact repository.
  • AWS: data preparation from de-identified cases, SageMaker Pipelines, fine-tuning (Llama / LoRA), evaluation and Model Registry, and an EKS/ECS staging replica monitored with CloudWatch.

Business use case

Keeps customer data and live inference inside the bank’s private cloud while using AWS for experimentation, training and evaluation. Mirrors the hybrid banking design worked on at Rackspace; production deployment was outside the demonstrated POC.

My role

Architecture design and documentation.