I run my own AI server rather than sending client or clinical data to a third party — a dedicated on-premises machine (dual Xeon processors, two GPUs giving roughly 48GB of usable memory) built to run serious open models like Llama 3.3 70B and DeepSeek R1 locally, with no cloud dependency.

In practice, this covers confidential audio transcription (5–30 hours a week, depending on the project) and private writing assistance on sensitive material — nothing leaves the building. Alongside the language models it runs Whisper for transcription and Nextcloud for storage, so the whole pipeline — audio in, transcript out, drafting, storage — stays on infrastructure I control end to end.

It's also what powers OpenClaw, an open-source personal AI agent, running on its own separate dedicated hardware.

Running this myself has shaped how I think about data ownership in practice — the same thinking I bring into AI adoption work with clients: what actually needs to leave your building, and what doesn't.