DEV Community

Paul Desai
Paul Desai

Posted on

I Run 107 Repos from a Mac Mini in Goa

#sovereignai #infrastructure #indiehacker #buildinginpublic

I Run 107 Repos from a Mac Mini in Goa

Author: Paul Desai
Date: 2026-02-18T17:00:00+05:30
Tags: sovereign-ai, infrastructure, indie-hacker, building-in-public, india, local-first

No AWS. No Vercel. No managed Kubernetes. Just a Mac Studio, a Mac mini, two phones, and a Cloudflare tunnel.

This is what sovereign AI infrastructure actually looks like.

The Stack

  • Mac mini M4 (24GB) — primary compute. Not even a Pro. Runs Ollama (14 models loaded), 15 services, 34 LaunchAgents, and Claude Code. The base model Mac mini. That's it.
  • Mac mini M1 (Red) — secondary node. Syncthing mesh, backup Ollama instance
  • Pixel 9 Pro XL — mobile command center. Tasker automation, Termux, MirrorBrain app, push-to-talk to Claude
  • OnePlus 15 — secondary phone. Same stack, doubles as hotspot when ISP goes down (today, actually)
  • Tailscale mesh — everything talks to everything, anywhere, encrypted
  • Cloudflare tunnel — public endpoints without opening ports

Total cloud spend: $0/month.

What Runs On It

  • Chetana — AI scam detection with Telegram + WhatsApp bots, serving real users
  • MirrorBrain — orchestration layer, model routing, identity-aware inference
  • Beacon — publishing platform with automated content pipeline
  • Cognitive Dashboard — real-time system health, device status, AI activity feed
  • Memory Bus — cross-session state persistence, agent handoffs, recall
  • 107 repositories — from governance layers to phone automation scripts

All of it self-healing. If a service dies, a LaunchAgent restarts it. If a phone disconnects, the mesh reroutes. If the internet goes down... well, I'm writing this on a phone hotspot.

Why Not Cloud?

Three reasons:

1. Sovereignty. My data, my models, my rules. No vendor can shut me off, change pricing, or inspect my traffic. When Anthropic has an outage, my local models keep running.

2. Economics. A Mac Studio costs what 4 months of GPU cloud would. It'll last 5 years. The math isn't close.

3. India. Internet here is good but not guaranteed. Power cuts happen. ISPs go down (like today). A sovereign stack degrades gracefully instead of going dark.

The Real Insight

The cloud isn't bad. It's a dependency. And dependencies are fine until they become single points of failure.

I don't argue against cloud. I argue for sovereignty — the ability to run your entire AI stack from hardware you own, in a country where you live, under rules you wrote.

If you can't run it without an internet connection, you don't own it.

What's Next

I'm open-sourcing pieces of this. The governance layer. The identity kernel. The scam detection engine. Not because I'm altruistic — because sovereign AI only works if others can verify it.

If you're building something similar, or you're in India and want to collaborate: activemirror.ai

Paul Desai is building MirrorOS — reflective AI infrastructure for autonomous systems. Based in Goa. Doesn't use the cloud.

Published via MirrorPublish

Top comments (0)