I've built a self-hosted sentiment analysis service running on Azure Kubernetes Service (AKS) using Ollama (phi3:mini). Yesterday I finally got it working in the Whizlabs sandbox. Today I completed the full Azure CI/CD pipeline. Cost control was a big factor for me: I'm running everything in my Visual Studio Professional subscription with only €45 in credits left!
What I built:
- Terraform for AKS cluster, Spot user nodepool, and static Public IP
- Multi-container architecture with a clean LLM adapter pattern
- Traefik Ingress Controller (with ARM64 considerations on the system nodepool)
- Cost optimization experiments (Spot vs on-demand, Dpsv6 system pool)
- Proper ConfigMap-driven prompts, health probes, and model pulling
- Cross-subscription deployment setup
- Reusable CI/CD pipeline with separate bootstrap + main stages
This project helped me connect my daily Azure DevOps work with the AZ-400 topics I'm studying. I learned a lot about real-world trade-offs, especially around cost optimization and infrastructure-as-code.
(Video demo coming soon)