Grounded in running systems, not takes.
Every article here links back to a live demo, a real policy file, or a real bug I found and fixed — not generic advice. Pre-launch, no client case studies exist yet; what follows is the actual engineering and governance reasoning behind what I've built.
Who's really running your AI?
Seven of nine enterprise systems I track each signed their own LLM-vendor deal in the last year — a system-by-system look at who anchored to which lab.
How a governed AI write-path actually works
Read, decide, approve, execute, audit, reject, rollback — with a real bug I caught building it.
NIST AI RMF, ISO 42001, EU AI Act: one mapping, not three checklists
Why the same evidence trail can support all three mapped frameworks, if it's architectural from the start.
An MCP integration still needs the same approval gates a custom API needs
What Model Context Protocol standardizes, and what it doesn't — with real code.
What actually drives Oracle Fusion Cloud ERP AI-agent readiness
A real, reproducible scoring model from use case, integration method, and existing governance controls.
What a real AI cost-governance ledger looks like
0 of 16 of my own sampled model calls settled at exactly $0; 15 touched billed credit — real numbers.
Why "auto-approve everything under $X" is an AP governance anti-pattern
Scope, spend tiers, and ERP validation as independent layers — plus a rollback bug I found.
Why router-watch and security-watch only propose — never apply
The real 12-day cross-machine auth gap that motivated security-watch, and why propose-and-approve is the whole point.
Oracle's own sanctioned path into EBS can't tell you which agent did what
Oracle's own documentation: HTTP Basic Auth only, a single shared service account for every call — verified against Oracle's own docs.
A vendor's governance module is the vendor's evidence about itself
ERP vendors are expected to ship their own agent-governance modules. What they can and can't evidence, with SAP's and Oracle's own documentation, and five questions to ask.
Prefer how the demos themselves were built? See the engineering writeups →