Case study 04 · KienlongBank · Dec 2024 – May 2026
KILOBA: a bank's AI platform, migrated live and run 30% cheaper
10+
AI services research → production
0
downtime, AWS OU migration
−30%
cloud compute cost
days → min
deployment cycle
// 01 the problem
The bank wanted one AI platform for every internal use case: customer-service chatbots, HR and training assistants, call scoring, computer vision, OCR. It had to run inside banking security constraints, serve thousands of staff, and ship ten-plus AI services in eighteen months. The infrastructure had grown up under a legacy AWS account structure; moving production into a properly governed Organizational Unit meant re-creating compute, networking and IAM while the bank kept operating. Cloud spend was growing with every new service.
// 02 architecture
One production OU, GitOps-deployed, observed end to end
Staff reach the platform through an APISIX gateway with Redis caching and an Open-WebUI front end with single sign-on. Workloads run on EKS inside a dedicated production OU, with RDS, S3 and Redis for state and Pinecone plus Neo4j for retrieval. LLM calls fan out to several providers through a routing layer with failover. GitLab CI builds, ArgoCD deploys; Prometheus, Grafana, Loki and Tempo observe; FinOps dashboards report monthly.
// 03 what I did and why
- Led the team as Tech Lead: technical direction, architecture review, delegation, and taking 10+ AI chatbots and services from research to production.
- Migrated production to a new AWS OU with zero downtime. Compute, VPC networking, security groups and IAM were rebuilt in the target OU and traffic moved over without an outage, leaving a cleaner IAM posture and clearer account boundaries.
- Onboarded GCP as a second cloud (project hierarchy, IAM, billing, Vertex AI / Gemini) so the platform could pick the best model per use case rather than one vendor.
- Standardised delivery. GitOps pipelines for the platform and partner integrations (payments, insurance, isolated QA), later packaged as two reusable CI templates (Kubernetes and Docker Swarm targets) that new projects adopt in an afternoon.
- Ran FinOps as a program: tagging, cost dashboards, billing alerts, right-sizing, storage lifecycle policies, monthly reports to management. Compute cost fell about 30% with no SLA change.
- Built the observability stack across all AI services, moving incident resolution from hours to minutes, and prototyped AIOps with Claude AI and MCP sub-agents for autonomous monitoring.
- Hardened the platform: Zero-Trust networking, Kubernetes RBAC, strict IAM governance, and an internal LLM gateway (LiteLLM) behind APISIX.
// 04 measured results
| Metric | Result | Source |
|---|---|---|
| AI services in production | 10+ chatbots and services | platform inventory |
| Microservices operated | 15+ on EKS | service specs |
| Production migration to new OU | zero downtime | migration record |
| Cloud compute cost | about 30% lower | FinOps reporting |
| Deployment cycle | days → minutes | pipeline history |
| Incident resolution | hours → minutes | observability rollout |
| Automation | 5+ n8n workflows, 10+ hours/week saved | workflow inventory |
| Reserved-instance coverage | 74% | cost report, Apr 2026 |
// 05 links
Bank code stays private. The CI/CD templates were packaged as a reusable bootstrap and can be published without bank-specific values.