The CodeLifter Blog

The flat line

The portal holds ten clonoSEQ results and every one of them is undetectable. The lab sent back the six earlier years, including one that went from 563 to 758,033. Both deliveries are complete, both are correct, and neither mentions that the other exists.

9 min read

A human clicks Download

Everything up to the download automates. The download does not. And the gap between what a browser reports and what is actually on disk is where a pipeline quietly starts lying to you.

6 min read

The record they give you is not your record

Epic will export your entire chart on request. The document it hands you first caps its results section at 200 entries, keeps the newest, drops the oldest, and says nothing about it.

I went looking for ten years of blood work and found seventeen

The census said twelve health systems. Opening the portals one by one, twelve became seven, the hardest paper chase evaporated, and the record turned out to start seven years earlier than I thought.

🧠 AI Engineering

The harness that only tested itself

I built a shared standards harness so Claude would follow the same rules across eleven repos. It had tests. The tests checked the harness, not the repos, and the first real audit came back with 34 failures. Then the standard started paying for itself.

August 8, 2026

11 min read

🤖 AI & Agents

The part no model fixes

I pointed an AI at seventeen years of my own medical records. It was confidently wrong three times before it was right once. No current AI untangles this alone.

August 4, 2026

🤖 AI & Agents

I kept failing to explain what an agentic team is. So I built one you can watch play football.

TouchDown isn't a fantasy football tool. It's a teaching artifact, a working agentic team built on a framework everyone already understands. The football team model for agent orchestration, in working code.

July 24, 2026

7 min read

The Census

Before extracting a single lab value, the project needed a census: who holds my blood work, from which years, and how each system expects me to get it back. The answer was about a dozen, and my inbox found them.

July 23, 2026

Ten years of blood draws, ten health systems, one spreadsheet

A decade of my blood work exists, scattered across ten health systems. I am going to collect all of it into one dataset I own. Part one: the plan.

July 21, 2026

6 min read

🤖 AI & Agents

Hundreds of lab reports, one conversation

A decade of blood draws scattered across four patient portals. Starting the project to collect them all took one paragraph typed into a chat box.

July 16, 2026

🛠️ Tools

The AI tool stack I actually use every day

For years the AI coding tools didn't impress me. Then last October someone said the word 'Cursor,' and I haven't come up for air since. The honest inventory, with benchmarks.

June 16, 2026

6 min read

💭 Ponderings

The work didn't go away. It moved.

Agentic coding didn't delete the work. It picked it up out of the middle of the pipeline and set it back down on the two ends. Same total, redistributed.

June 13, 2026

2 min read

💭 Ponderings

Hallucination is just imagination showing up at the wrong moment.

We curse hallucination as the cardinal sin of LLMs. It's the same property we praise when the model writes a test we didn't think of. The bug and the feature are one thing, pointed two directions.

June 11, 2026

2 min read

Before I got here, there was MajorBoost

The founder story I've been avoiding telling for years. A medical AI startup, a few hundred lessons, and where it all came apart. It's the reason I'm finally writing about agents.

June 9, 2026

7 min read

💭 Ponderings

The most underrated skill in 2026 is writing a prompt other people will use.

Most prompts get treated like a Slack message to your past self. The ones that actually work get treated like code somebody else is going to read.

May 28, 2026

2 min read