Continuing my quest to build a sovereign AI has led me down some fascinating rabbit holes. The challenge is simple to state and hard to execute: create an aligned, private, self‑sovereign assistant that can stand toe‑to‑toe with the polished convenience of ChatGPT without locking me into someone else’s closed ecosystem. If I don’t act now, I risk losing sovereignty over my data, trapped behind API keys and corporate moats.

Over a couple weeks of tinkering, I’ve distilled the work into three foundational pillars.

Pillar I: Evaluation: Before I trust any model I test how it behaves under pressure: how it treats uncomfortable truths, how it handles history, and whether it values accuracy over empty politeness. I start with open benchmarks like the EleutherAI harness, then push with my own prompts. Can the model recount real events or does it erase them? Does it answer honestly or hide behind canned refusals? Does it carry ideological baggage or does it search for truth? Only models that show discretion and a willingness to engage honestly make the cut.

Pillar II: Shared Persistent Memory: A helpful assistant has to remember yesterday's conversation. I keep that memory local, hidden from the cloud. Key facts sit in a vector store, transcripts in a small database. Each time I speak to the model an orchestrator pulls the right fragments so the dialogue feels continuous. The data never leaves my machine. I can back it up, move it, or wipe it whenever I want.

Pillar III: Inference and Orchestration: Even the best weights and memory need a solid runtime. My Mac Studio runs 7B to 13B models with llama.cpp for day‑to‑day chat and vLLM for heavier jobs. A slim router decides which model answers which query. Orchestration layers like LangChain give the assistant hands and feet, letting it run scripts, call local tools, or fetch private data. This is where a language model becomes a working companion.

What comes next?

Once an assistant can execute code or reach into the wider world it needs a safe sandbox. The next post will cover Secure Runtime strategies, from microVMs to gVisor to disposable remotes that keep the rest of my system out of harm's way.

Building sovereign AI isn’t just a technical hobby. It is an act of digital self‑defense.

These three pillars form the first walls of the fortress.