An Operator's AI Lab: Series Overview

For the last few months I’ve been obsessed with what I can actually build with AI agents. Not chat with. Build. Hands-on, trying to solve real problems.

I didn’t go down this road to crank out demos or slides. I went down it to run parts of my own life: the family finances, the house, my health, the home network itself. Real systems, with my real data, that I depend on. Personal production, not a sandbox.

Most writing about AI agents is pretty generic and mostly un-actionable, and it fails in two opposite directions. Enterprise whitepapers that never touch a keyboard. “Hello world” wrapper scripts you wouldn’t miss if they broke. I simply can’t read about another inbox triage agent (and yes, I built one of those too). The interesting part is the gap neither one describes: pointing agents at your own messy data, somewhere being wrong actually costs you something. It’s harder than the scripts and realer than the whitepapers. And the thing that genuinely surprised me is how much I could just ask the agent to do — and it did it.

That’s the lab. Here’s why I’m writing it up. None of these builds is world-changing on its own — it’s a house, a gym log, a pile of PDFs. But I run a data company, and every wall I hit at home turned out to be one my customers are hitting right now: what’s actually worth automating, what data you’d never send to the cloud, which model is worth paying for, how far to trust an agent with the keys. I’d rather learn that with my own messy data than off a slide.

So each piece below is a build — and what it taught me about doing this for real. Each piece has a corresponding build-notes article that walks through step-by-step how the solution was built. The Getting Started build notes below provide an overview of how to get started with the stack.

Here’s the series:

1. Turning My House into a Thermal Battery

Under California’s NEM 3.0 solar rules, the excess power my panels dump back to the grid in the afternoon sells for pennies — right before peak rates hit at 4pm. So I asked my agent: what if we stored that energy in the house itself? It precools the place on free solar, then coasts through peak. It even spec’d and flashed firmware onto a $26 chip to make a 10-year-old furnace smart without wrecking its efficiency. This is the version of AI nobody demos: an agent that acts on the physical world and pays for itself, instead of writing you a poem. Read the Article →

2. A Private AI Fitness Coach

I’m not a biohacker. I’m a busy CEO who cycles through travel, steak dinners, and long gym absences, and every calorie app I’ve tried made me want to throw my phone across the room. So I built a coach that runs private, local models to ingest ten years of Apple health data, blood panels, and my DNA — data I’d never hand to a cloud service — and made it dead simple to use (“I ate a BLT and fries,” and it does the rest). Keeping that data local is what made it good. Privacy wasn’t the tax I paid on the AI; it’s the reason the AI was worth having. Read the Article →

3. A Private CFO for the Family Books

Years of financial history scattered across a dozen institutions, in hundreds of PDF statements nobody wants to read. My agent wrote the parsers, loaded it into a private database, and paired the numbers with plain-text catalogs of the wills, trusts, and policies — all under strict cost guardrails so it never burned a dollar it didn’t have to. Now I can audit fee drag, catch subscription creep, and run “what if we moved states” in seconds. Nobody’s real data is clean and in one place — mine sure wasn’t. Getting an agent to make sense of the mess, without running up a cloud bill doing it, is most of the actual job. Read the Article →

4. Multi-Model, Cheap Hardware, No Excuses

You don’t need a $4,000 Mac Studio or a $400/mo enterprise plan. Every system in this series runs on an M2 Mac Mini with 8GB of RAM and two $20/mo consumer subscriptions. The agent installed and wired the whole thing itself: a cloud model for hard reasoning, a second one as an adversarial code reviewer, and a free local model for anything private. Turns out the trick is matching each task to the cheapest model that can handle it — the biggest brain is usually overkill. Read the Article →

5. The Remote Control Plane

The plumbing under all of it. How I run this stack headless at home and drive it from an iPad anywhere in the world — over a zero-trust network, behind hardware security keys, with sessions that survive a dropped connection mid-task. You don’t hand an agent the keys to anything that matters until this part is solid. Get it right and everything else becomes safe to try; get it wrong and you shouldn’t be building any of the rest. Read the Article →

I’d genuinely welcome feedback, questions, and takes on what’s missing.

I was a mediocre — but, I’d argue, innovative — software developer early in my career. I joke that I wrote such good code they introduced me to product management. This project got me back into building for the first time in years, even if I’m not the one typing the code anymore. And it’s landed me squarely in the same problems I spend my day job on: multi-model, cost, privacy, security, and getting real value out of agents — beyond asking ChatGPT to clean up an email nobody bothered to write themselves.