Posts

This is Phase 3 of the hybrid AI inference platform. In the previous post, I connected a Mac Mini running Ollama to a GCP VM through Tailscale. That works for direct connectivity, but it is a single tunnel with no redundancy and no dynamic routing. This phase adds the enterprise-grade networking layer: HA VPN with BGP. What I tried first (and why it failed) My original plan was to build HA VPN directly between GCP and my Mac Mini, using Tailscale as the underlay network. The Mac Mini’s Tailscale IP is 100.80.121.94, and I configured it as the peer endpoint for the VPN tunnels in Terraform. ...

This post covers Phase 1 and Phase 2 of my hybrid AI inference platform. Phase 1 gets local inference running on a Mac Mini. Phase 2 builds the cloud side with Terraform and connects the two through a private network. By the end, a GCP VM in us-central1 is calling an AI model on my Mac Mini in Lincoln, UK, through an encrypted tunnel. Part 1: Local Inference Why local inference matters Cloud AI inference is powerful but expensive. Every API call to Vertex AI or OpenAI costs money. For straightforward tasks like summarisation, Q&A, and code explanation, a smaller model running on your own hardware handles the job without the bill. ...

I could have used Medium. I could have used Hashnode or Dev.to. But I’m a cloud engineer, and the way I publish my writing should reflect that. This blog is a Git repository. Every post is a Markdown file. When I push to main, a GitHub Actions pipeline builds the site with Hugo and deploys it to GitHub Pages. My custom domain is managed through Cloudflare. The whole thing is infrastructure-as-code, version-controlled, and automated. ...