Writing

Learning to ask for help, then learning to be asked Editor's pick

What it took to stop trying to do everything alone, and what I owe the people who wouldn't let me.

Handing the AI work to other teams

Standing up the company's AI function solo: workflow interviews, a license-holder champion network, a shared agent library, and where it's heading.

Scoring the ADA agent against WAVE

Checking the agent's fixes against a real accessibility tool built on the WebAIM Million. Average score went from 6.3 to 7+ out of 10; contrast errors were the last thing holding it back.

Building a local AI agent from scratch with Ollama and OpenClaw

Skill modules, memory architecture, and why I stopped relying on the cloud.

Owning a legacy app across four systems

What it means to be the subject matter expert on software older than you'd like.

How I built a harmonic pattern detector in Pine Script

Ratio math, pattern memory, and staying under TradingView's token limit.

An ADA compliance agent that won't touch code without asking

Why the human-approval step is the feature.

Teaching the ADA agent to write its own patch files

A senior dev's demo suggestion, SVN patch files I'd never made before, and a "no patch needed" bug.

Teaching an agent your codebase's unwritten SQL rules

Convention extraction, schema verification loops, and where the time actually goes.

Testing the limits of vibe coding with Claude

Free tier, paid chatbot, artifacts, connectors, Claude Code: what changed the outcome wasn't the subscription tier.

A few years of watching AI go from confidently wrong to actually useful

College ChatGPT, a failed vibe-coding attempt, model-switching, and the vibe-coding economy I still go back and forth on.

How I ended up building the company's first MCP server

A CEO's idea, a 40-to-50-hour manual process, and an MCP server I pitched before I'd built one.

Redesigning our EMR's UI without a frontend team

Thirty years without a UI update, no dedicated frontend team, and what AI closed the gap on.

Letting the ADA agent scan an entire codebase on its own

The per-file loop was the bottleneck. A multi-agent full-codebase scan is projected to cut the ADA estimate from 50-60 hours a system to about 5, pending a full-system test.

Rebuilding the internal site's frontend in three days

A second senior dev asked for the demo on the real site: reusable CSS classes, color tokenization, a frontend rebuilt in three days.

Two MCP servers I built for my own dev workflow

Questions about the database and the codebase that used to mean opening SSMS or digging through SVN, answered in natural language instead.

Turning the UI rebuild into a design system

The internal site rebuild is still in testing, and it's already led to the next thing: a documented design system for the team to build from.

Scaling our MCP servers with AWS AgentCore, Bedrock, and Lambda

A meeting with three AWS engineers I hadn't planned to lead, and a crash course in AgentCore, Bedrock, and Lambda to scale MCP servers past my laptop.