Notes
Field notes from building Bob — a local AI agent for SQL Server self-healing — and the surrounding territory: monitoring, regulated data, agent architecture, and the art of letting an LLM touch production without setting it on fire.
The Onboarding Sprint: How to Operate a 200-Table Database You've Never Seen Before
You inherit a database with no documentation, 200 tables, 400 stored procedures, and a senior DBA who left in March. The traditional onboarding takes two weeks. With an LLM in the loop, it is two days. Here is the playbook.
Anomaly Detection: When to Reach for Statistics, When to Reach for AI
Threshold alerts catch the bug you predicted. Statistical methods catch the drift you didn't. LLMs add the explanation that makes the alert actually useful at 2 AM. Mixing the three buys you the kind of monitoring that catches plan regressions before they become incidents.
Index Recommendations: Don't Trust the DMV. Don't Trust the AI. Use Them Together.
SQL Server's missing-index DMV recommends every index it thinks would help a query — and is right about half the time and catastrophic the other half. The AI alone is no better. The combined workflow gives you a smaller set of better-targeted index changes.
Stop Using AI to Write Your SQL. Use It to Read Your SQL.
Most DBAs first reach for AI as a code generator. That is the wrong direction. The real value of an LLM in a database role is as a reader, not a writer — of execution plans, stored procedures, schemas, and the T-SQL you were already going to ship.
The 9-Second Suicide: How ‘Vibe Coding’ Deleted a Startup and the Architect’s Guide to AI Guardrails
Nine seconds. One overscoped API token. A startup's database, gone. The PocketOS incident is a case study in why "vibe coding" without the Principle of Least Privilege is the architectural equivalent of leaving the master keys in the door.
MCP and the Three-Layer Agent: How to Let an LLM Touch Production Without Setting It on Fire
An LLM that can diagnose but not act is a smart alarm. One that can act without snapshotting first is a liability. The boundary between the two is mostly architecture: sensors that observe, reasoning that recommends, actuators that mutate — with a snapshot-before-apply rule that fails closed.
The Birth of Bob: An AI Agent for SQL Server That Runs on a Single Home GPU. The book is up.
Eighteen months. One RTX 3090. A book about why local inference is the right answer for regulated DBAs — and why the same DBAs drowning in 200-page health reports are exactly who Bob is built for. The book is up.
SQL Server Projects: Extend, Don't Decorate
AI should not be a shiny coat of paint on top of your SQL Server estate. It should be a fundamental extension of your operational discipline. The fundamentals still matter; AI just makes the people who do them faster.
AI as the On-Call DBA's Second Brain
PagerDuty rings at 2 AM. You are staring at a wait-stat combination you have never seen in this particular configuration. The DBA's loneliest moment is exactly where AI delivers its most measurable value — as a runbook reader, not as an actuator.
AI Assistant vs. Human Receptionist: Is It Time to Upgrade Your Front Desk?
The front desk is your first point of contact. As 2026 progresses, more business owners are weighing 24/7 AI phone assistants against full-time receptionists. The honest answer is usually a hybrid — here's the breakdown.
Schema Archaeology: Letting an LLM Read 20-Year-Old Stored Procedures Faster Than You Can
Every DBA inherits a database written by a person who has since left, retired, or been promoted out of caring. Schema archaeology used to be a multi-day exercise. With a language model reading procedures faster than you can, it is one afternoon.
Moving from Broadcast to Conversation: The AI for DBAs LinkedIn Group Is Live
The newsletter response was bigger than expected. To turn this from a one-way broadcast into an actual community, the AI for DBAs LinkedIn group is now live, alongside the Skool community where the deeper technical work happens.
When Cloud LLMs Aren't an Option: Local Inference for Regulated Database Workloads
Sending a query plan from a HIPAA, NERC CIP, or GDPR-bound environment to OpenAI is a six-month vendor review, not an API call. Local inference skips that conversation entirely & works on a consumer GPU.
The Report Problem: Why Your Weekly SQL Server Health Check Is a Tombstone
I once buried a joke about cat photos on page 87 of a 214-page SQL Server health-check report. Nobody mentioned it. That is how I knew the report had become a tombstone, and why DBAs need variance detection rather than more pages.