Dipilatoautomations

☕️ I don't check my email anymore.

I just say:

"Hey Jarvis, give me my morning brief."

While I make coffee, my AI reads my email, calendar, weather, and reminders. Then it tells me the only things I actually need to know.

The private stuff stays on a computer in my house. Big jobs go to the cloud—but only when I choose.

It feels like having a real-life Jarvis.

Comment "JARVIS" and I'll send you my setup.

One sentence books the meeting, sends the email, and pulls my numbers — across calendar, email, texts, Obsidian and my wiki. Ten tabs collapsed into one command.

Tailscale connects my phone to my home computer over a private network, so I can run my AI assistant from anywhere — with nothing exposed to the internet.

My AI assistant routes every task across three brains — a local model first, then mid-tier, then frontier — so most work never leaves my computer. Private by default, expensive only when it has to be.

$0/mo, zero rate limits, and 100% private code on your machine. The catch: 12 GB has a ceiling and setup is an afternoon, not one click.

For some builders, that trade is worth it.

You know the moment: the coding agent reads the whole repo and you just wait. The fix is not only faster writing. It is getting the prefill read done in one breath.

What part of your agent loop feels slowest?

This sounds fake until you understand mixture-of-experts models. For any token, only a few experts fire. REAP prunes the dead weight and keeps the router plus the workers.

That is how a 28B model fits in 12 GB.

This is the local AI build that surprised me: a used RTX 3060, 12 GB VRAM, no API, no subscription. Stock was about 205 tok/s. Tuned prefill hit 1,142 tok/s on the same card.

Would you try this on a used GPU?