← Blog
02 — Topic
Artificial Intelligence
11 posts on this topic.
- 3 min
You Don't Need a Bigger Model. You Need Better Retrieval.
A 4-billion-parameter model on a CPU-only server went from 200-second waits and off-topic citations to useful answers by fixing retrieval, keeping the model loaded and gating it with an eval harness.AI · RAG - 3 min
Five ServiceNow API Calls That Return Success and Store Nothing
Five ServiceNow Table API behaviors that report success while storing or returning nothing — currency fields, ampersands in references, bad choice values, class filters and partial access — and the habit that catches them.ServiceNow · CMDB - 4 min
A Dry Run Is Not Permission
Guarded writes to a production system: why a clean dry run proves nothing about write access, and the preview, intent, window, canary and read-back pattern that makes AI-assisted updates safe.ServiceNow · AI - 2 min
Introducing ea-brain: A Starter for an AI-Maintained Architecture Knowledge Base
ea-brain is a free, MIT-licensed starter for an AI-maintained architecture knowledge base: the contract, resume state, rules and skills from this series.Enterprise Architecture · AI - 2 min
Pushed Is Not Deployed
Ten releases logged as deployed while the server sat on an old version. Why a push is not evidence of a deploy, and three rules that keep the two apart.DevOps · AI - 3 min
Every App Needs Two Capabilities
Map every application to business capabilities for what it serves and exactly one technical capability for what it is. The second lens is what makes rationalization work.Enterprise Architecture · Capability Modeling - 5 min
Two Tiers: Spend the Expensive Model Only on Judgment
Split long-running AI work by how much reasoning it needs: a strong model for judgment and schema, a cheaper model running bounded skills, and one rule that decides between them.AI · Claude - 6 min
Sessions Forget: Designing Resume State for AI Agents
AI sessions lose everything that lives only in context. A resume page, a ruled-out list, and fixed close and boot rituals keep long-running AI work intact across days and machines.AI · Claude - 5 min
Never Let the Generator Grade Its Own Work
How a separate, skeptical AI reviewer — one that assumes the work is wrong, runs the checks, and never fixes its own findings — made an AI-maintained knowledge base trustworthy.AI · Claude - 5 min
Green Tests, Broken System: 4 Ways a Test Suite Lies
Four ways a passing test suite hid real defects while I built AI tooling — fixtures built on assumptions, rules that never fire, checks that stop looking, and gates that defend the bug.Testing · AI - 6 min
Building an AI Second Brain for Enterprise Architecture
What worked in building an AI-maintained knowledge base for an enterprise architecture practice: three layers, resume rituals, a skeptical reviewer and cite-don't-restate.AI · Enterprise Architecture