Ground Floor

I run this project from a home office. None of it is theoretical. The results live or die on this specific setup: two M5 Max MacBook Pros connected over Thunderbolt 5 RDMA, 256 GB of unified memory, 16 TB of storage, ~3 µs latency between nodes. That's the same hardware I write about on this site.

Why this work

For a while I kept seeing the same conversation repeat. A solo physician asks about AI documentation tools. The products people recommend, all of them excellent, require an internet connection, a vendor agreement, and some number of calls with a compliance consultant. That physician, who runs a single-provider practice with two staff and no IT department, quietly closes the browser tab.

The tools that exist for large health systems don't translate well to small practices. Tools built for small practices often handle the compliance question with a checkbox and a promise. There's a third option that almost nobody is covering in a systematic way: air-gapped AI. Run the model locally, on hardware you already understand, and don't create a third-party data flow in the first place.

That's the gap Ground Floor is trying to close. Not with a product. This is an education project, not a company, and the way it closes the gap is with clear documentation of what actually works.

The format

Each week I pick one thing to test: a model, a hardware tier, a task, an industry. I document the setup, the results, the failure modes, and what you'd need to replicate it. The verdict answers the question most people actually have: is this worth my time?

The writeups live here. Discussion happens on LinkedIn. This site is the reference layer.

What I'm not

I'm not a lawyer, a compliance consultant, a physician, a financial advisor, or a licensed professional in any of the industries I write about, and I want to be clear about that because I think it matters. What I am is someone who understands air-gapped AI infrastructure and wants to make that knowledge accessible to people who don't have time to piece it together from technical papers and Discord servers.

Every experiment result is a technical finding. The compliance implications for your practice are a separate question that requires someone who knows your specific situation. My Scope & Disclaimers page says this more carefully.

The lab setup

Two M5 Max MacBook Pros on a Thunderbolt 5 RDMA cluster, 256 GB unified memory, 16 TB, ~800 GB/s peak bandwidth. It's more hardware than most people will buy. But it lets me test the full spectrum: throttle the config down to match a $799 Mac mini, or open it up for 70B context workloads. When I say something works at entry level, it ran at entry level.

Contact

Best way to reach me is on LinkedIn or X. I read everything. If you're a practitioner in a regulated industry and want to share a use case, challenge a finding, or suggest an experiment, I want to hear it. The experiments get better when they're grounded in real workflows.