Sovereign AI engineering log
A self-hosted AI stack built and documented in public. Every article comes from a real system, a real error, or a real fix. Nothing is speculative.
For people running, scoping, or considering self-hosted LLMs on DGX Spark, Strix Halo, multi-3090, or similar hardware.
Who is cipherfox, why sovereign AI, and what the stack looks like right now.
Read about the projectAll articles ranked by EEAT quality score: expertise, experience, authority, trust.
View content quality
Emanator is a full Buffer clone for Nostr with Rails, database, and background jobs. Our Python script is 492 lines, zero dependencies. Continuation of the Anti-Slop article: what we borrow, what we don't, and why Sovereign AI often means less code.
Read article
I asked Qwen3.6 to redesign a stats component on sovgrid.org. It did, with real unique readers from NSM, git-based edit history, collapsable layout, and a dozen bug fixes along the way. The cost was eighteen commits, eight debugging iterations, and a growing sense that I was stress-testing my local LLM at the edge of its useful range for a feature whose value to me as operator was genuinely questionable.
Read article
Qwen3.6-35B built a working SvelteKit 5 + D3.js app from a single prompt. The build passes. The features ship. But the architecture is improvised, the code is duplicated, and it took eight debugging iterations to get here. Here's what the gap between 'works' and 'good' actually costs.
Read article
A self-hosted stack still hits two or three tasks where a frontier model wins. Buying that access from Anthropic means a KYC account and a card. ppq.ai is the other door: an OpenAI-compatible proxy to Claude, GPT and others, paid per query over Bitcoin Lightning, no account. Here is what it is good for, where it betrays the sovereign premise, and exactly how I wired it as the fallback behind local Qwen.
Read article
Gemma-4-31B in NVIDIA's NVFP4 format fits a single DGX Spark and is a strong reasoner. But on Blackwell sm_121 the default FP4 kernel path is broken, and a dense 31B is bandwidth-bound at around 4 tok/s no matter what you do. I measured the baseline, the Marlin fix, and the honest conclusion: the real speedup is a model swap, not a flag.
Read article