Story

From the build log

We build in public. Each note is a short, honest look at what we shipped, what broke, and what we learned.

Three-panel comic. A builder pins a card to a big public roadmap board with Shipped, Building now, and Planned lanes, a diverse community cheers and raises coins to vote, and a lane lights up Funded under a blue constellation. A banner reads: Build in public, vote with dollars.
Building in public

Building in public

Every story here, every number and honest miss, I put out in public on purpose. This is why I build in the open, and how you can help shape what comes next.
Spotlight comic. In a cozy night reading nook, an open glowing book pours friendly code into the air that becomes a blue star constellation and a tiny running app, with a small gold desktop computer glowing beside it. A banner reads: Read it, run it yours.
Books

A book you can read and run

I write books you can run, not just read. The code works and every page is online. Here is why open, reproducible books beat theory behind a paywall.
Three-panel comic in a warm workshop. A tired builder glues the same parts by hand, a tall tower of repeated glue-tax parts looms over them, then they open a tidy blue toolbox of ready-made modules. A banner reads: Pay down the glue tax.
fieldkit

The glue tax, and the kit that pays it

Every AI project, I kept rewriting the same boring plumbing. I named it the glue tax. So I built fieldkit, an open kit of 12 tested parts, to stop paying it.
Emblem-style comic with four glowing panels: an open book labeled Books, a friendly app window labeled Software, an AI brain chip labeled Models, and a small gold desktop computer labeled Devices, tied together by a blue star constellation. A banner reads: Four things you control.
AI Native

Four things you actually control

The AI divide feels out of your hands. It is not. There are four things you can own instead of rent. Playbooks, software, models, and a device of your own.
Editorial comic. A crowd is stuck behind a wall while a few people walk through a door marked Learning into bright light, where a small gold desktop AI glows.
AI Native

Why most teams get nothing from AI

Companies poured 30 to 40 billion dollars into AI and 95% got nothing back. The reason is not cost or tech. It is learning. That gap is why Orionfold exists.
A builder smiles at a night desk under a starry sky as a small gold computer folds a glowing blue constellation down onto the desk. A banner reads: Fold it down. Open, private, yours.
Founder

Why I folded Orionfold

After nearly nine years at Amazon, I left to start Orionfold. This is the twenty-five-year story behind that choice, and what the name means.
Three-panel painted comic. A builder looks overwhelmed at a shelf full of glowing AI model boxes and a tangle of terminal windows, then sketches a single dashboard beside a glowing gold desktop computer, then sits back happy at one screen with live dials, a leaderboard, and two models racing, under a blue star constellation. A banner reads One cockpit for every model.
Tools

The cockpit for my models

I had a shelf of models on one desktop and no way to drive them. In fifteen hours I built a cockpit to run, compare, and score them, all on that desk.
Spotlight comic. A small friendly AI model wears a graduation cap and holds a badge that says My Field, standing on a small gold desktop computer while the builder claps.
Fine-tuning

Teaching a small model my field

I taught a 3B model my own field with 231 question-answer pairs. It stopped refusing and started answering, in my voice. Small and tuned beat big and general.
Three-panel comic. A builder runs one command, gets a first word in 52 milliseconds, and a small gold desktop AI glows with a blue constellation of icons.
DGX Spark

My first model on a desktop

I ran my first model on a small computer on my desk. 52 milliseconds to the first word, no cloud, no per-use bill. It felt like a local function, not a service.
Two-panel comic. On the left a builder is stuck choosing from a wall of model names. On the right the same builder ships happily with one gold desktop AI.
Local AI

Access first, models second

On day one with my desktop AI machine I did not pick a model. I set up how I reach it. Models change every six months. Good access lasts for years.
Four-panel comic. A builder writes three sentences at breakfast, the plan fans out into tasks, an app fills the screen by noon, then the builder leans back by the finished app.
Spec-driven

A spec at breakfast, an app by lunch

With Kiro I turned three sentences into a 24,000-line app, about four hours of my time. The spec did the heavy lifting, and speed became a skill.
Three-panel comic. The same AI agent meets a developer at a terminal, a manager at a dashboard, and a person by voice.
Agents

One agent, three faces

The same AI agent should meet a developer in the terminal, a manager in a dashboard, and a phone user by voice. One brain, many doors.
Comic. A builder works at a desk inside a warm glowing dome at night while data stays safely in the room.
Privacy

Keeping my data in the room

I built an analyst AI that never sends my work to the cloud. Privacy is not a setting you flip at the end. It is a choice you make in the design.
Four-panel comic. A builder plays chess against a glowing AI at night, looking relaxed at first, then leaning in focused and making the winning move. A banner reads: Vibe coding is not passive.
Vibe coding

Vibe coding is not passive

After my AI-built weekend I thought coding got easy. Andrew Ng named what I felt. Vibe coding is not passive, it is engineering done fast.
Three-panel comic of a builder at a home desk turning a weekend idea into a finished app with AI help.
Vibe coding

I built a real product in a weekend

Over one holiday weekend I built Web Memo, a working browser tool, about 2,400 lines of code. AI wrote most of it. My sense of my own ceiling moved.
Splash comic. A giant glowing price tag tears from 100k down to 10k in a grand hall, while a golden AI brain and a small gold desktop computer glow nearby. A banner reads: Frontier reasoning at a tenth of the cost.
Open models

Frontier reasoning at a tenth of the cost

In early 2025 a model called DeepSeek did frontier-level reasoning for about 95% less. And it was small enough to run on my own machine.
Two-panel comic at a sunset stadium. In 2023 a runner labeled CLOSED leads a smaller runner labeled OPEN by a wide margin. In 2024 the two hit the finish tape side by side under a blue star constellation. A banner reads: The year the gap closed.
Open models

The year the gap closed

In late 2024 I watched the AI tool stack collapse into the model itself. The first sign that one person could build what used to take a team.
Subscribe to stories

Get each new story in your inbox.

Short, honest notes from building Orionfold in public. What we shipped, what broke, and what we learned. No spam.