Sovereign AI Architecture

I am writing an email to Lex Fridman and want to shorten and tighten this: Big fan Lex, ever since your machine learning lectures at ocw.mit.edu, I feel a shift in AI coming and I just wanted to see what you think. Sovereign AI is artificial intelligence that runs on infrastructure you fully own and control, data, models, and compute, without dependence on external providers. Every day it seems like local models become better and better, Qwen3.5 is a game changer and I think is a tipping point where local options are becoming more and more realistic. But what is missing from simply running Ollama and chatting with a .gguf is everything you get from corporate API models, the governance, the guardrails, the ownership and control of data and compute. That is what I have been addressing. Building a sovereign stack where I use llama.cpp or vLLM for inference of any model I want, then I stack my data, I took 2 years of my writing, channeling a dead friend through Reddit posts, then construct a knowledge graph out of the posts and all my other written material along with a vector db. On top of that I stack a harness, such as ByteDance’s DeerFlow 2.0 and that really allows me to get the type of responses you typically pay an API for. On top of that I created my own architecture I call Dynamic Persona Mixture of Experts RAG. It allows more observability, granular control and hallucination mitigation. On top of that I can create any frontend I want or use Telegram, etc. e716f91233a1567c9b3ebc23fbd6fa3ee8ac7a3b22f1bc44bfce6d9595b9bbc0.png If you are curious about this architecture in more detail or just curious about what I have been teaching myself, I created a knowledge graph with my vibe coded blog: https://danielkliewer.com The key is that it all runs locally on my laptop, not needing the internet or to pay for API access or experience the API rate limits or not having control over my data or only getting a triaged version of compute. I really do think this is the future of AI. Especially if we see political backlash to AI, that is, concerns about environmental impact and centralized control. I am actually from Austin and have lived here my entire life. I have been interested in AI since I first started working on Amazon Mechanical Turk in 2007, it is my main hobby.

:::writing{variant="email" id="48392" subject="Thoughts on Sovereign AI"} Big fan of your work, Lex—especially your MIT OCW machine learning lectures. I’ve been following the direction AI is heading and wanted to share a concept and get your perspective. “Sovereign AI” refers to systems where the data, models, and compute all run on infrastructure you fully own and control, without reliance on external providers. It feels like we’re approaching a tipping point where local models are becoming genuinely viable. Recent advances like Qwen3.5 make it increasingly realistic to run capable systems locally. What’s still missing, in my view, is the ecosystem that corporate APIs provide—governance, guardrails, and integrated tooling around data and deployment. I’ve been working on building a “sovereign stack” to address this. It combines local inference (llama.cpp / vLLM), personal data layers (knowledge graphs + vector databases built from years of writing), and orchestration frameworks to approximate the capabilities people typically rely on via APIs. On top of that, I’ve been experimenting with an architecture I call Dynamic Persona Mixture-of-Experts RAG, aimed at improving observability, control, and reducing hallucinations. The entire system runs locally without requiring external APIs. I’ve documented parts of this on my site if you’re interested: https://danielkliewer.com I’d be curious to hear your thoughts on where you see this space heading, especially as local models continue to improve and questions around centralization, cost, and control become more prominent. Appreciate your time, and thank you for all the work you’ve put out over the years. Best, Daniel :::

now explain dynamic persona mixture of experts RAG in a sentence

Dynamic Persona Mixture-of-Experts RAG is an architecture that routes queries through multiple specialized persona-conditioned retrieval-and-generation pipelines, dynamically selecting and blending the most relevant “expert” perspectives over your indexed knowledge to produce more controlled, context-aware, and consistent responses.

no base it on danielkliewer.com

Dynamic Persona Mixture-of-Experts RAG is a locally run system that builds multiple persona-specific “experts” from the writing and knowledge graph derived from danielkliewer.com, then dynamically selects and combines the most relevant personas with retrieval from vector and graph memory to generate responses that reflect different styles, perspectives, and contextual expertise.