Notes from the field.
Long-form writing on private AI deployment, on-premises infrastructure, compliance, and the engineering decisions behind every system I ship.
Apple Silicon as an inference node: M4 Max & M3 Ultra, honest digits
Benchmarks for 70B models on M4 Max and M3 Ultra. Why Apple is betting on local inference - and what the token economics tell us about the future.
Public AI assistants in higher education: the GDPR exposure most institutions have not assessed
When staff paste student work into a public AI assistant - ChatGPT, Claude, Gemini, whichever - the institution becomes the controller for a processor it never contracted. A walk through GDPR Articles 5, 28, 32 and 35, the rulings already issued, and the architectural fix that does not require banning AI.
A private LLM for a research lab: notes from a 14-day rollout
Field notes from a 14-day on-premises private LLM deployment for a 28-person genomics lab: Mac Studio M3 Ultra running Llama 3.3 70B via Ollama, AnythingLLM RAG over 2,400 unpublished documents, pfSense deny-by-default egress, and zero outbound bytes after handover - with the hardware trade-offs, GDPR and grant-compliance framing, and the three things that broke.
Private RAG for contract review: a law firm case study (Dubai/London)
How a Dubai/London law firm cut contract review time by 73% with a private Llama 3.1 70B RAG running on-prem over 12k binding documents - audit-grade citations, NDA-safe, no cloud.
Hand-written referrals to structured EHR records: a clinic NER case study (UAE)
A UAE clinic group turned hand-written referral letters into structured EHR records in 4 seconds with 99.2% NER precision - on a single on-prem A6000, never touching the public internet.
Why 40% of AI Projects Get Canceled - And the Five Decisions That Separate the Rest
Gartner says more than 40% of AI agent projects will be canceled by the end of 2027. The reason is almost never the technology. Here is what actually goes wrong - and a practical framework for not ending up in that pile.
How I Built an Autonomous AI Content Engine for a Crypto Media Company
A crypto media company needed to cover a 24/7 market with a finite editorial team. Here is how I built an autonomous pipeline that went from detecting events to publishing finished articles - and why the hardest part had nothing to do with AI generation.
One field note a month. No marketing.
If you run sensitive data and want a heads-up when I publish, drop your email. No tracking, no upsells, no AI roundup of someone else's blog.