What we're learning while building AI systems that have to work on Monday morning — not just in the demo.
Ox Alpha was Z.ai's GLM-5.3-Flash. We ran one BFCL benchmark on it: 100 samples, too few to separate it from our local 35B model. Cost: under four cents.
We put a four-node Turing Pi cluster next to our DGX Spark to give it eyes, off-box backups, and GPU senses behind one LiteLLM front door. The build, the honest numbers, and why the 'GPU node' runs on the CPU today.
Self-hosted AI image generation on our DGX Spark made brew-space's launch hero — a reproducible, commercially-licensed 45 KB WebP at zero marginal cost.
Our LLM benchmark said 0.84 while agent runs failed. The culprit: speculative decoding corrupting vLLM output — found with a $0 determinism probe.
HumanEval, BFCL and MBPP on a DGX Spark: a 33B coding specialist vs a 35B generalist, the parallel tool-call zero we chased to the model, and the caveats.
We added HumanEval, MBPP and BFCL to our DGX Spark benchmark lane, caught the BFCL bug that rewards never calling a tool, and tried free Ox Alpha.
NVIDIA's Nemotron 3.5 Lightning benchmarked on DGX Spark 24 hours after release — plus the Ollama regression and runaway-reasoning failure the harness caught.
We built a local AI image + video studio on a DGX Spark in a day. Real GB10 benchmarks, the license traps verification caught, and the quality ceiling.
A big release landed for our self-hosted AI agent. We contained it first — and found two systemd directives that don't do what they look like.
Our unlaunched website served 14,000 requests to nobody. How to read pre-launch traffic, pick cookieless web analytics for $0, and the trap we found.
DGX Spark benchmarks for gpt-oss-120B and Laguna-S-2.1 — first independent numbers — plus the silent GPU bug that corrupted our published results.
Real DGX Spark local LLM benchmarks: a 35B MoE at ~100 tok/s, why dense 70Bs are a trap on unified memory — with corrected serving-stack numbers.
Adversarial AI review, applied to our own pitch deck: 29 agents attacked every claim. All six moat claims survived only after being rewritten honest.
We built an AI chief of staff meant to have one tool. AI agent security bit us: 'allowed tools' didn't restrict anything. Here's the fix.
We built a Slack-to-Copilot Studio relay in one night. Five things the docs don't tell you — auth, billing, and one SDK bug that matters.
Multi-agent panels, an adversarial trademark sweep, a licensing-audited pipeline: our mascot, made in an afternoon for $0 in new software spend.
OpenTofu is a drop-in replacement for Terraform — except for one sharp edge in the lock file that our CI found before we did.
How we moved our own site from a template builder to versioned infrastructure on Azure with Cloudflare DNS — practicing what we preach.
The gap between a great demo and a dependable product is wider for AI than for any software we've built before. Here's how we close it.
When retrieval-augmented generation disappoints, teams reach for a bigger model. Nine times out of ten, the fix is upstream — in the data.