Everything I've learned building a sovereign AI stack, in public: the proof
(case studies), the failure log (dead ends written up because they were the
expensive part), and the interactive field guide.
Problem β solution β outcome, compressed to the result. Built and measured
on my own stack: where a number is from a real run I say so, and where
something didn't work I say that too. Each links to the full story.
I built a quantum ML lab. The honest answer: quantum mostly ties or loses today
Reproduced a published quantum advantage (0.900 vs 0.730), then erased it three ways: a fairly tuned classical model beat it, dequantization matched it, device noise took the rest. Eleven experiments, two real QPUs, zero manufactured wins.
One product photo β a marketplace-ready 3D model + a compliance report
Five real SKUs on one RTX 5090: three passed clean, one refused, one flagged with the exact miss (703mm reconstructed vs 650mm listed). The gate stopping bad assets is the product.
One shot list β an Unreal-ready previs pack in 15 minutes
Six props (4 photos + 2 text briefs): five delivered in 15m12s warm, one refused with its reason shown. One idempotent Unreal import script, Nanite on heroes only. Tested in Unreal Engine 5.8.
vLLM + Nemotron with routing-enforced privacy: the client profile refuses to fall back to cloud, failing loudly instead of degrading silently. Operated daily at up to 268 tok/s, 180K context.
Auditing a private codebase for vulnerabilities, fully air-gapped
Hybrid system beats the LLM alone, 0.97 vs 0.90, and the cross-file taint engine hits 1.00 precision where a per-file LLM gets 0.25. Honest headline: the fine-tune added ~0 detection capability, and measuring that caught 4 of my own eval bugs.
A talking-head avatar from a photo and a voice sample, three quality tiers
For a 34.5s narration, measured: MuseTalk 78s, LatentSync 242s, Hallo2 1197s. Honest ceiling: social-grade; broadcast close-ups still go to a cloud path, and I say so up front.
Clean removal in 180 seconds, no ghost, background reconstructed. On a worst-case clip (glassware, low light) it leaves a residue, and clients hear that up front with a free sample frame.
A vision LLM writes the scene constraints itself before any diffusion runs: 203 seconds end to end, unattended, 16 mask targets including 13 dependent reflections and shadows. The known hard case (a liquid) did not disappear, and that boundary is named.
Panel detection, cloned character voices, narrator on captions, 2.5D parallax camera. Runs end to end; in active hardening on real series pages, hard problems documented as theyβre solved.
New posts as the work happens.
Looking for something specific? Browse by topic β
(LLM serving, GPU/VRAM, quantum, creative AI, audio, opsβ¦)
A GTK switch for one GPU shared by seven AI stacks GTK4 mode switcher for an RTX 5090 shared across seven local AI stacks: the idle-vs-busy detection problem, an SM-util false positive from reverse-PRIME graphics, and why the app is a front-end only.
One YAML file to rule my estate How PROJECT-STATUS.yaml became the single source of truth for 22 projects, and the fail-open trap where a stray quote made the dashboard serve convincing lies.
My security auditor is 80 percent assembled, not invented How BHAIRAVA shipped as a real one-person security product: 80% existing measured components, 20% new connective tissue, and Stage V honest failures including 0/12 recall on real Django CVEs.
A local 3D asset foundry with two license lanes TRELLIS.2 vs Hunyuan3D on one RTX 5090, QuadRemesher tracing-stop when Python launches it, and why the VLM judge failed a good asset the pixel metrics cleared.
Planning a robotics stack before buying a single servo VISHWAKARMA P0: Isaac Sim quarantined for sm_120 Blackwell bugs, the target drone discontinued, and GR00T weights carrying a non-commercial license the Apache badge did not show.
A fitness AI that refuses to guess ADASHINO plans training with a deterministic solver -- the LLM is the conversational shell, not the source of the numbers. No demo video, and a four-week dogfood gate before any claims ship.
How many qubits fit in 32 GB? I simulated until the GPU refused A hybrid quantum-classical net on MNIST 0/1 lands at 99.5% β just behind logistic regression (99.6%), the honest loss. Plus the memory-wall sweep: the statevector fits to 28 qubits, then the card taps out.
I let an LLM agent design quantum circuits in a sandbox for 10 cycles An LLM proposed VQC variants, each run in a locked no-network sandbox, scored vs a 0.978 baseline. It explored, hit a collapse, recovered, and edged the baseline by +0.0044 β within noise. The honesty and the sandbox are the story.
The quantum kernel that got worse as I added qubits A fidelity quantum kernel loses to a classical RBF-SVM on both datasets β and its off-diagonal similarities concentrate 64Γ as qubits grow, the kernel version of the barren plateau. An honest negative result.
My code auditor scored 1.00 on fixtures and 0.00 in the wild 13 real Python web apps, 276 findings, zero true positives. The five rule bugs real code exposed, the β90% fix pass that still found nothing, and the LLM oracle that coin-flips on the hard case.
Sovereignty as routing, not policy How to enforce 'client data never reaches a cloud API' with LiteLLM routing profiles β including a chain that deliberately errors instead of degrading silently.
The forensic bridge: measure the scene before diffusion touches it A vision LLM does a three-pass forensic scan of the footage β heroes, reflections, atmospheric fields β and writes the SAM3 mask prompts and fill constraints itself. Full pipeline in 203 seconds, autonomy proven, and the one target it honestly couldn't remove.
What I designed in May vs what shipped in June I generated an architecture diagram for my local AI stack. It was gorgeous. Within a month, seven things in it were wrong β and I couldn't diff the PNG to find what changed.
Moving Docker's data root doesn't move containerd I moved Docker to the big disk and /var kept filling up anyway β because Docker and containerd are two different daemons with two different stores.
I benchmarked an AI avatar pipeline end to end Ten local audio and avatar pipelines on one RTX 5090 β TTS, voice clone, transcription, lip-sync, music, dubbing β measured wall-clock and VRAM for each, including the two pipelines that stayed blocked and the close-ups that still go to cloud.
Six hours in TensorRT-LLM so you don't have to I picked NVIDIA's 'official' inference path for a multimodal model and spent six hours hitting walls, ending on one AutoDeploy can't fix. The lesson: 'theoretically compatible' is not 'someone benchmarked this.'
The SAM3 mask that crashed VACE A segmenter was handing my video editor a single still-frame mask shaped [1,H,W] when the model wanted one mask per frame, [N,H,W] β a tensor-dimension mismatch that looked like everything except what it was.
Nightly wheels are a depreciating asset Rebuilding a one-month-old audio-stack Dockerfile failed with xformers ResolutionImpossible on cu130, because the nightly index it pinned had moved on. Pin proven stable, never pip-install at runtime.
A 32GB GPU is a budget, not a suggestion How to fit a 30B multimodal LLM at 180K context on one RTX 5090: move the display to iGPU, never co-load models, and pick NVFP4 over FP8.