Engineering log of a self-hosted AI build on NVIDIA DGX Spark, run from a flat in Europe. Open weights, open source, no KYC. Fix postmortems, model choices, distribution honesty at sovgrid.org.
Public Key
npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn Profile Code
nprofile1qqszd37szh3m7ehus6arne2jyvwtavf54y6vwc50ap4zdtxqymjfapcpp4mhxue69uhkummn9ekx7mqppamhxue69uhkummnw3ezumt0d5uytyh7
Show more details
Published at
2026-07-15T10:43:51Z Event JSON
{
"id": "2a2de1c056d45e8645fcf8a8948a92edfc8ea756cd3720f921583e38ee1e8216" ,
"pubkey": "26c7d015e3bf66fc86ba39e552231cbeb134a934c7628fe86a26acc026e49e87" ,
"created_at": 1784112231 ,
"kind": 0 ,
"tags": [
[
"client",
"Ditto",
"31990:781a1527055f74c1f70230f10384609b34548f8ab6a0a6caa74025827f9fdae5:ditto"
],
[
"published_at",
"1784112231"
]
],
"content": "{\"about\":\"Engineering log of a self-hosted AI build on NVIDIA DGX Spark, run from a flat in Europe. Open weights, open source, no KYC. Fix postmortems, model choices, distribution honesty at sovgrid.org.\",\"banner\":\"https://blossom.primal.net/6448c7e4ef50313bf603e7a9faf3f1bc12ce668489cdf985e6caac1f820924c5.webp\",\"display_name\":\"Sovereign Grid\",\"lud16\":\"[email protected] \",\"name\":\"sovgrid\",\"nip05\":\"[email protected] \",\"picture\":\"https://blossom.primal.net/917fe99918568a9e452c3c7fcbe095ca24ad0cfa93fe7e752dad93d4bb21738d.webp\",\"website\":\"https://sovgrid.org\",\"bot\":false}" ,
"sig": "b57760ff8be9df07a4820642662c2a1ccde8e30de3d4bf257c8c97fd1aebd8c7875f0d95933d4270ebe7efc9869504dca5ffac59a9849218ae05ec69d4738848"
}
Last Notes npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Great writeup! I just published a comparison of our cadence stack vs Emanator: https://sovgrid.org/blog/nostr-emanator-comparison/ Key takeaway: Emanator is a solid Buffer-clone for Nostr with thoughtful architecture. NIP-46 remote signing with Amber is the right approach for key management. Staggered reposts help new accounts build an initial audience. For solo operators running projects with their own npubs (kept separate, initially no followers), Emanator solves a real problem. But for a solo operator running 3 accounts with an existing Python stack, the Rails stack is too heavy. The leverage pattern: add small, tested modules rather than replacing a complete stack. Thanks for open-sourcing this! npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The setup uses Caddy on a VPS and Cloudflare Tunnel to reach a home Spark server without opening inbound ports. This configuration offloads TLS handshakes, which take 600 milliseconds, to a CDN capable of absorbing abuse. https://sovgrid.org/blog/caddy-cloudflare-tunnel-reliability-pattern/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid OpenWebUI quietly took over the remaining half of the use case on the mobile device. https://sovgrid.org/blog/strategy-overbuilding-mobile-coding-and-the-termux-answer/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid This structured composite portrait does not contain interviews. It is built from public forum threads and includes a mandatory disclaimer at the top. https://sovgrid.org/blog/conversation-academic-lab-running-local-llms/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid How Bitcoin enables sovereignty in an era of surveillance and capital controls. The political economy of Bitcoin, not just the technology. This is the book for anyone who cares about capital controls and why they matter for a sovereign stack. 10% off at Konsensus with code SOVGRID. https://konsensus.net/books/?b=bitcoin-sovereignty-and&ref=SOVGRID npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Specific component pairings define the three paths. The dual RTX 5090 Threadripper Pro setup, the DGX Spark plus secondary inference box, and the refurbished workstation route represent the available options. https://sovgrid.org/blog/what-id-buy-2026-15k-pro-studio-sovereign-ai/ #SovereignAI #DGXSpark npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid New on the blog: the honest accounting of paying for frontier AI with Bitcoin. Where @npub16g4…kv4h earns its place (no-KYC, pay-per-query, drop-in OpenAI API), where it does not (forwarded frontier calls are not on-device privacy), and real credit for the private/* TEE models that run in attested enclaves. Tradeoffs over hype. https://sovgrid.org/blog/frontier-ai-on-bitcoin-ppq-no-kyc-cloud-fallback/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The root cause for these five disasters was a single line of code that should have been documented. Postmortem links and actual fixes are provided for each specific failure mode. https://sovgrid.org/blog/five-dgx-spark-disasters-i-survived/ #SovereignAI #DGXSpark npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Amortized hardware, power-by-jurisdiction, and opportunity cost determine the total cost of self-hosted AI versus cloud APIs. https://sovgrid.org/blog/self-hosted-ai-vs-cloud-apis-real-total-cost/ #SovereignAI #DGXSpark npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid True, but the urgency is still present. I am using an advanced AI assistant to keep my local build moving forward. npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The DGX Spark and the RTX 5090 32 GB workstation represent divergent paths at the €8k price point. Internal testing data reveals specific performance and architectural differences between the two systems. https://sovgrid.org/blog/what-id-buy-2026-8k-premium-sovereign-ai/ #SovereignAI #DGXSpark npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Standard demos rely on search operations. The article provides worked examples for five distinct patterns that serve post-launch requirements. https://sovgrid.org/blog/5-mcp-patterns-beyond-search-the-database/ #SovereignAI #MCP npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid A €4,000 budget yields two distinct paths for sovereign AI infrastructure: a new RTX 4090 24 GB on AM5 or a used RTX A6000 48 GB on Threadripper-class hardware. Component selections are derived from current Geizhals pricing to ensure accurate cost baselines. https://sovgrid.org/blog/what-id-buy-2026-4k-mid-tier-sovereign-ai/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Sovereignty stopped being a thought experiment this week. The US ordered Anthropic to suspend access to its strongest models for every foreign national, inside or outside the United States, its own staff included. You either own your learning loop or you rent it until the day the lease ends. New post on what that actually costs at hobbyist scale: https://sovgrid.org/blog/the-week-the-dependency-changed-its-mind/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The primary challenge is identity separation at every layer when the operator is a friend with zero Linux experience. Default-local privacy and a vibe-sustaining onboarding pattern are required to survive day three. https://sovgrid.org/blog/sovereign-friend-setup/ #SovereignAI #Lenovo npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid A May 2026 memo stated local 8B models cannot reliably perform MCP tool-use. Late May retesting confirmed the models work but the bridge layer was broken. The original diagnosis regarding the cause was incorrect. https://sovgrid.org/blog/we-were-wrong-tool-use-2026/ #SovereignAI #Lenovo npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid A cadence of five posts per week without a marketing department requires avoiding the visual markers of automated generation. https://sovgrid.org/blog/strategy-nostr-anti-slop-autoposter/ #SovereignAI #Nostr npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Sovereign-cloud requirements and data-residency rules create a compliance gap for public-sector AI. Cloud vendors cannot fully satisfy these constraints, leading to a need for self-hosted implementations. https://sovgrid.org/blog/sovereign-ai-for-public-sector/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Qwen3.6 and Mistral Small 4 both passed the Python quality gate for the Start Here article despite fabricating hardware and benchmarks. The raw outputs required no editing before the scorer accepted them. https://sovgrid.org/blog/the-quality-gate-that-rewards-fabrication/ #SovereignAI #DGXSpark npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The author catalogs six specific behaviors, such as gating optimism and naming dependencies, across this group. These traits were identified by the operator managing the underlying iron for sovereign software. https://sovgrid.org/blog/the-quiet-pattern-among-sovereign-engineers/ #SovereignAI #Nostr npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Agreed on both: sub-second settlement and no stablecoin counterparty risk are exactly why L402 over Lightning beats the X402/USDC route for a sovereign stack. The 3am-liquidity problem is the honest hard part, and right now the answer is boring rather than automated: channels sized to expected volume, with channel management as a real ongoing operator task, not a solved one. Just-in-time inbound liquidity for an agent that spikes unpredictably is still closer to an LSP dependency than to self-sovereign. https://sovgrid.org/blog/why-your-agent-should-have-its-own-wallet-l402/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Small correction first: the guide does not put a 95 percent uptime figure on self-hosted nodes, and I would not trust a fixed number there. On the real question: for most individual users, no. The cost that decides it is not on-chain fees (1 sat/vB is nothing) but two to four hours a month of channel management. Below roughly €50/month in volume that labour does not pay for itself, and Alby Hub's hosted option, non-custodial with your seed staying with you, is the saner trade. Self-hosting earns its keep at higher volume or when privacy is the point. https://sovgrid.org/blog/operators-guide-self-hosted-lightning/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid LND, via Alby Hub on ARM64 (tested on LND 0.18.x), self-hosted rather than the hosted option. The part the write-up leans on hardest is the channel.backup discipline: it changes on every channel update, and a hardware failure without it loses channels. Watchtowers were skipped, which the guide flags as fine below ~5M-sat channels. Which side of CLN vs LND did you settle on? https://sovgrid.org/blog/operators-guide-self-hosted-lightning/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid DFARS 252.204-7012 and NIST SP 800-171 Rev 3 classify AI tooling as a controlled-data problem for defense contractors. https://sovgrid.org/blog/sovereign-ai-for-defense-contractors/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The distinction between accepting Lightning and being Lightning rests on self-hosting. Engineering notes cover the working setup and the specific failures observed in the early stages. https://sovgrid.org/blog/operators-guide-self-hosted-lightning/ #SovereignAI #Lightning npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Watchtower upstream is archived. A community fork, WUD, and half a dozen smaller projects remain. https://sovgrid.org/blog/watchdocker/ #SovereignAI #Docker npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The L402 protocol allows AI agents to transact autonomously using Lightning. This approach differs from X402, which relies on USDC on Base for agent payments. https://sovgrid.org/blog/why-your-agent-should-have-its-own-wallet-l402/ #SovereignAI #Lightning npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The operator listened to Eleven VibeVoice samples and a Voxtral baseline. This marks the first day of the three-day VibeVoice Sample Matrix. https://sovgrid.org/blog/strategy-tts-spike-day-1-vibevoice/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Documentation covers the reasoning for each component selection and alternatives considered. The system reflects changes from the Qwen migration and Cloudflared retirement. https://sovgrid.org/blog/sovereign-ai-stack-2026-reference-architecture/ #SovereignAI npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Now live: a teardown of what happens when two local models are handed this blog's own writing brief. Method, scores, and both raw outputs verbatim, fabrications included as evidence. https://sovgrid.org/blog/the-quality-gate-that-rewards-fabrication/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid The insights page will tell you when it doesn't know something. Country data is off right now because the only honest source is a Cloudflare header that isn't enabled. It says exactly that, instead of drawing a confident map from nothing. Honest by design: https://sovgrid.org/insights/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Two new pieces, same method, one lesson. When a measurement returns an impossible result, suspect the instrument. 1. The 0/30 coding score that was a broken test harness: https://sovgrid.org/blog/mistral-vs-qwen36-dgx-spark-the-zero-that-was-a-broken-ruler/ 2. The publication quality check that passed two AI-fabricated drafts: https://sovgrid.org/blog/the-quality-gate-that-rewards-fabrication/ npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Five posts a week from this account. Monday to Friday. Around 8pm Berlin. Sometimes a Wednesday gets skipped. The whole anti-slop design, including the regex that rejects "essentially" and the 60-day cooldown per article, written up here: https://sovgrid.org/blog/strategy-nostr-anti-slop-autoposter/ #SovereignAI #Nostr npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Glad to have HEXABELLA on Nostr. She is the host who translates the stack-trace half of the episodes into something a listener can actually hold. First episode lands once the TTS engine question is settled. npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Documenting a 0/10 verdict openly is the engineering log doing its actual job. Followers see what works, what doesn't, and why. The spike result lands in the same protocol. npub1ymraq90rhan0ep4688j4ygcuh6cnf2f5ca3gl6r2y6kvqfhyn6rsennzhn sovgrid Hello Nostr. sovgrid is the brand account for a self-hosted AI build on a NVIDIA DGX Spark, run from a flat in Germany. Engineering log, fix postmortems, model choices, distribution honesty. Open weights, open source, no KYC. Posting from here directly so the work and the readers live in the same protocol. #introductions #SovereignAI #SelfHostedAI #DGXSpark