New write-up: A Second Brain for a Local Model, and the Two Bugs That Made It Useless First.

Sovereign AI Blog

A Second Brain for a Local Model, and the Two Bugs That Made It Useless First | Sovereign AI Blog

Giving a local 8B model persistent memory and retrieval good enough to replace a cloud assistant for daily coding. The architecture is mem0 plus a ...

#SovereignAI #MCP

I replaced hidden complexity with info-buttons on every metric and a glossary that explains every acronym. This shift from admin tool to teaching surface fixed the adoption problem.

Sovereign AI Blog

Dashboard As Learning-Cockpit, Not Admin-Tool | Sovereign AI Blog

I had a 600-line dashboard that worked technically and went unopened socially. Rebuilding it as a teaching surface changed everything. This post is...

#SovereignAI #Lenovo

The first time I tried to route MCP through a custom dashboard, the whole RAG pipeline collapsed under its own weight.

Sovereign AI Blog

24 Hours Setting Up a Lenovo Legion Pro 7 Gen 10 As a Sovereign-AI Companion Box | Sovereign AI Blog

Honest minute-by-minute log of building a friend

#SovereignAI #Lenovo

Small honest thing: my stats page now says "country data unavailable" with the reason, instead of guessing. To get real country data I'd have to flip a Cloudflare header on. I'm genuinely unsure I want it. Is a number worth the dependency?

Sovereign AI Blog

Insights | Sovereign AI Blog

Live content quality dashboard: weighted quality scores, image KPIs, and reader engagement, powered entirely by a local AI pipeline.

Follow-up to this. The grid runs Qwen3.6 on the Spark now, Mistral is the fallback that turned out to be the only one of the two that can actually read an image. Two rulers, both nearly lied to me. Full story:

Sovereign AI Blog

Mistral vs Qwen3.6 on DGX Spark: the 0/30 That Was a Broken Ruler | Sovereign AI Blog

I almost published

View quoted note →

Question for people running self-hosted models for real work: how do you fact-gate AI-written technical content before it ships? My regex scorer rewards confident fabrication. A human read is the only thing that caught it. Is there a better layer, or is "a human who knows the machine" just the actual answer?

The thing nobody tells you when you build a quality check for AI text: a fabricated benchmark and a real one are the same shape on the page. The script can measure structure. It cannot measure truth. That part is still on me, and I almost forgot it.

Ran an experiment on my own blog. Gave two local models the same brief for the intro article I still owe this site, took their raw output, and fed it straight to my quality gate untouched. That gate is an automated check a draft has to clear, not the thing that ships it. I still make the final call, for now anyway. Both passed the check. Both had quietly invented hardware I don't own and a maintenance routine I never run. The gate didn't catch it. It gave them points for it, because a fake version number looks exactly like a real one to a script that only counts. Wrote the whole thing up, both raw outputs included as evidence:

Sovereign AI Blog

The Quality Gate That Rewards Fabrication: I Had Qwen and Mistral Write This Blog | Sovereign AI Blog

This blog gates every article behind one Python scorer before it publishes. I gave Qwen3.6 and Mistral Small 4 the same brief, the Start Here hub a...

Still learning this. Mostly learning that my own tools lie to me politely.

Eight Voxtral fix articles. Three weeks of patches. Then I sat down and actually listened to Episode 1 V6. Zero out of ten. Short turns hallucinate "ähm, ähm" filler. Long turns flatten into staccato. There is no tuning sweet spot in the middle. The open checkpoint has no path to release quality. I am done patching it. Next spike: VibeVoice / Higgs Audio v2 / IndexTTS-2.

Sovereign AI Blog

Voxtral Capped at 3/10: Picking the Next Open TTS | Sovereign AI Blog

Eight engineering fixes deep, three weeks of patches, two failure modes on the same engine. The Voxtral open checkpoint has no path to release-qual...

#SovereignAI #TTS #DGXSpark #Voxtral

This is exactly why I started: sovgrid.org View quoted note →

Self-custody Bitcoin, no-KYC, V4V. The discipline maps onto AI almost verbatim. Cloud inference is sats on an exchange. Local inference is cold storage. Same trust-removal, different layer. I just wrote a Start Here for folks asking "ok, what's the how":

Sovereign AI Blog

Self-Hosted AI: Start Here | Sovereign AI Blog

The honest decision tree for taking a local LLM stack seriously. Hardware tradeoffs I actually made, the inference engine I picked and why, the par...

@BTC Sessions your channel covers the why. Thank you for 10y awesome BTC content! My blog is the "what broke and how I fixed it" layer underneath your brand new sovereign AI channel. #Bitcoin #sovereignAI p.s. my blog is also queryable via MCP at mcp.sovgrid.org/self-hosted-ai no signup, if your agent does that. View quoted note →

1st challenge done: Two days from localhost to production. NVIDIA DGX Spark, Mistral Small 4 119B running local, no cloud APIs. Three failure modes that almost killed it (unified-memory OOM on Docker restart, IPv6-blocked HF downloads, SGLang quirks on ARM Blackwell) all the fixes + full reproducibility checklist:

Sovereign AI Blog

Two Days From Localhost to Production: Building a Hybrid Sovereign AI Site | Sovereign AI Blog

A two-day build log from localhost to a sovereign hybrid AI site. Three failure modes, exact fixes, and the reproducibility checklist most cloud gu...

Zap if it helped. 🙏 Reply if I missed something. Got a sovereign-stack story of your own? Drop it below. Repost for anyone tinkering on local LLMs.

Hi, Cipherfox here :-) A few weeks ago I was a complete AI noob. Now I'm tinkering on a sovereign stack: Mistral 119B running local on a DGX Spark, zero cloud. Learning in public. Follow along and help me level up. #introductions #sovereignAI

Happy Birthday, Satoshi Nakamoto!