Analysis you can use on Monday.
Long-form, evergreen pieces built from primary sources and public disclosures. These are decision guides for technical leaders: what the technology actually costs, where it breaks, and how to choose between real options. Unlike the daily brief, they are not tied to a single news cycle.
Startups & Venture
AI vendor due diligence: 25 questions to ask first
Model provenance, data retention, evaluations, SLA, and pricing cliffs. The questions that separate a serviceable vendor from an expensive dependency.
AI & Machine Learning
AI agents in production: seven failure modes and fixes
Tool misuse, injection, loops, stale context, permission creep, cost blowups, silent regressions. What each looks like in production, and the guardrail that catches it.
AI & Machine Learning
RAG or long context: choosing a retrieval design
Million-token windows did not retire retrieval. Cost, latency, and where models actually attend decide the architecture, usually a hybrid of both.
AI & Machine Learning
Open weights or closed APIs: a CTO framework
The choice is not ideology. Four constraints (data, customization, compliance, volume) decide it, and self-hosting math is harsher than most decks admit.
Cloud & DevSecOps
Kubernetes cost traps the platform bill hides
Cluster compute is the smallest line. Control planes, observability, NAT, idle requests, and platform labor add up to more, and here is the checklist.
Cloud & DevSecOps
The GPU capacity math: how many you really need
Sizing AI infrastructure starts from tokens per second and cache memory, not FLOPS. A worked example shows why peak sizing doubles the bill.
Cloud & DevSecOps
Cloud repatriation math: when leaving AWS pays
Repatriation cases live or die on staff cost, not egress. A five-line break-even model, a worked estate, and the conditions that say stay.
AI & Machine Learning
Why AI token prices fall while your AI bill rises
Unit prices keep falling and invoices keep climbing. The gap comes down to tokens per task, batching, and cache reuse, and it is modelable.