My AI Lives in Two Places at Once

The Brain/Body Split

I spent months trying to cram everything into one place. First I tried all-cloud: rent a beefy VPS, run the model remotely, pipe results back to my Mac. It worked — until I looked at the monthly bill. GPU-backed cloud instances are not cheap, and every query felt like watching coins tick down. Plus, I had zero control over the model version, the inference pipeline, or what happened to my prompts on someone else’s metal.

Then I tried all-local. Keep everything on the MacBook — the model, the automation, the webhooks, the scheduled jobs. That worked too, until I closed the lid. A laptop on Wi-Fi cannot reliably run 24/7. Discord bots go silent at 2 AM. Cron pipelines miss their window. And the moment I left the house, my entire AI stack went dark.

So I stopped choosing. I split it.

The brain lives here. On my desk, inside a 2026 MacBook Pro M5 Max with 128 GB of unified memory and 4 TB of storage. It runs DeepSeek locally — the ds4 model, roughly 284B total parameters with about 13B active per token. Inference is fast, private, and costs me exactly $0 in monthly compute. No API keys, no rate limits, no “sorry, the service is overloaded.”

The body lives in the cloud. On a VPS — a small, always-on Linux box running Hermes middleware and n8n automation. It handles the always-on tasks: receiving webhooks, orchestrating workflows, running scheduled pipelines, and talking to the outside world so my sleeping laptop doesn’t have to.

The brain thinks. The body acts. And they talk to each other over a thin, encrypted bridge.

The Core Rule

Before I show you the wiring, here are the hard specs so nothing gets lost in translation:

Local engine: Apple Silicon Mac (2026 MacBook Pro M5 Max, 128GB RAM, 4TB storage). Model: DeepSeek ds4 (~284B total params, ~13B active per token). Ongoing compute cost: $0/month.

VPS host: TODO: VPS provider. Running Hermes + n8n for always-on middleware and automation.

Everything I publish here is true to these constraints. If I name a specific host or URL later, it will be filled in only when finalized. Until then, you will see clear TODO markers.

The Roadmap Ahead

This post is the starting line. Here is what comes next in the series:

The Local Setup — Configuring the MacBook + DeepSeek ds4 engine for private, offline inference.
The VPS Middleware — Standing up Hermes + n8n on a cloud host for always-on automation.
Wiring the Two Together — Mistakes I made bridging local inference with remote webhooks, and the latency lessons learned.
Real-world Projects — The Dinner Assistant and the Email Aggregator: two practical workflows you can clone.
Closing the Loop — The automated publishing pipeline (veda-dev) that ties it all back to this very website.

Each post will include downloadable blueprints — n8n JSON exports and shell scripts — so you can adapt the patterns to your own stack.

Closing

If you build AI on your own terms — local inference, cloud automation, zero subscriptions — you are in the right place. The project files and workflow exports will live right here on the site. Explore them at vedabyte.ai.