Field Notes
Self-hosted AI, written down.
Technical essays from two years of running a private AI stack in production — the TCO math, the failure modes, the compliance reality, and what the vendors leave out.
Named or Assumed: The First Question I Ask of Any Architecture Document
The weak point of most architecture dossiers is the sentence that says security is 'handled by the platform.' Here's the one distinction that separates auditable inheritance from a gap in disguise.
My AI Router Was Spending $20 a Month Asking Itself "Are You Okay?"
A default health-check interval was quietly burning 43,000 paid API calls a month. Finding it took a cost dashboard; fixing it took one line. The real lesson is about defaults.
€8K Server vs €60K/year OpenAI Enterprise: When Each One Actually Wins
The unvarnished TCO math of self-hosted AI versus an OpenAI Enterprise contract — including the operator costs both sides of the argument leave out.
GDPR Article 28 for AI Vendors: What OpenAI's DPA Actually Says (And What It Doesn't)
A verifiable breakdown of what OpenAI's Enterprise DPA guarantees under GDPR Article 28 — and the compliance gaps that stay on your shoulders no matter what you sign.
The 3 Hidden Failure Modes of Self-Hosted LLMs (When You Scale Past 25 Users)
KV cache fragmentation, the context-window I/O trap, and kernel scheduler thrashing — the three failure modes that crash self-hosted LLM stacks at scale, and how to fix them.
Sovereign OCR: When Scanned PDFs Eat Your AI Pipeline (And Your Compliance Posture)
You self-hosted the LLM but pipe documents through a SaaS OCR API? The most sensitive part of your workflow is now the leakiest. How to close the hole properly.
vLLM vs llama.cpp vs Ollama at 25 Users: What the Published Benchmarks Actually Show
Single-user local inference is a solved problem. Production LLM serving at 25 concurrent users is a different engineering domain — and the published benchmarks all agree on the shape of the failure.