Writing.
Notes on building products and training models — in public, as I go.
The Glass Box: Building AI You Can Actually Watch Think
I built a layer that shows you an AI's decision the moment it acts — before it acts — and right now it's watching every command on my machine. Here's why agents you can audit are the thing that unlocks agents you can trust, and how I built one in a weekend.
Read article →I Gave My AI Idle Time and It Started Drafting Revenue
When my AI has nothing in front of it, it doesn't idle. It pulls up the revenue streams I actually run and drafts the next piece of money work — spread evenly, never the same well twice, and never past a hard firewall. It proposes. I still decide.
Read article →The Reflective Layer: Four Minds in a Markdown File
I gave my AI a layer that doesn't do work — it reflects on the work. Four small agents named Self, Conscience, Voice, and Awareness, each written in plain markdown, each asking a different question about what the system just did. Here's why a mind needs more than one voice.
Read article →The Gate Every Money Decision Has to Pass
Before any money decision happens in my system, it runs through a gate that encodes two rules I refuse to break — don't gouge, and never reprice a loyal client. It's not a vibe in a prompt. It's deterministic code, and right now it shows me its verdict before I act.
Read article →The Day My AI Started Fixing Itself
For months my AI could watch itself break and do nothing about it. Then I closed the loop — gave it a narrow list of repairs it's allowed to make to its own body, and a heartbeat that heals the parts that rot. Here's what changed, and where I drew the line.
Read article →My AI Safety Tool Publishes Its Own Failures Every Time It Runs
Most AI tools show you a green checkmark and hide the misses. Mine prints the five attacks it can't catch, by name, every single time you run the benchmark. That list is the most honest thing in the whole project — and the most important.
Read article →I Taught My AI to Never Raise a Loyal Client's Price
I underprice a client on purpose, because he took a chance on me when nobody else would. So I wrote that loyalty into my AI as a hard rule — because an optimizing agent will quietly undo your principles the second you stop watching.
Read article →Safety Is Not Ethics — and Confusing Them Is How AI Hurts People
A force-push that erases your teammate's work is dangerous but not unethical. Gouging a customer is unethical but perfectly safe. Most AI guardrails only check one of these — which means they wave the other one right through.
Read article →I Built an AI Safety Tool That Blocks Nothing. On Purpose.
My AI safety layer watches every command an agent runs, decides whether to refuse it — and then lets it happen anyway. That's not a bug. It's the only way I'd ever trust it to say no.
Read article →AI for Law Firms: Start With Intake, Not a Robot Lawyer
Every pitch aimed at law firms right now is some version of an AI that practices law. That's the wrong place to start and the riskiest one. The boring, unglamorous front door — intake — is where AI actually pays for itself in a small firm, with none of the malpractice exposure.
Read article →Buy the Asset, Not the Subscription
The entire AI industry wants your business on a meter. Per seat, per token, per month, forever. But the most valuable AI a small business can have isn't rented from anybody — it's built once, owned outright, and gets smarter on your own data. Here's how to tell the asset from the bill.
Read article →I Built a GUI So I'd Never SSH Into a GPU Again
Fine-tuning a model is mostly a tangle of SSH sessions, rented GPUs, and shell scripts only the person who wrote them can run. I got tired of being that person. So I built Forge — a desktop app that turns a training run into a few clicks, and turned my own toolchain into a product.
Read article →I Built the Marketplace for a Truck, Not a Desk
Contractors don't run their business from an office chair. They run it from the cab of a truck between job sites, with one hand on the wheel and a phone going off in the cupholder. So when I built FairTradeWorker, the app wasn't the afterthought. It was the whole thing.
Read article →Measure Twice, Deploy Once
Every carpenter learns the same rule before they're trusted with a saw — measure twice, cut once, because you can't un-cut a board. It turns out that's the same rule that keeps you from blowing up a production database. The job site taught me more about shipping software than any tutorial did.
Read article →The Billable Hour Meets the Language Model
A whole profession bills by the hour, and a technology just arrived that collapses certain hours to seconds. That sounds like a threat to the business model. Looked at straight, it's the opposite — but only for the firms that get honest about which of their hours were ever worth charging for.
Read article →What I'd Tell a Contractor Before They Buy Software
Every construction software company is circling small contractors with the same pitch and the same monthly bill. I came out of the trade and now I build this stuff, so let me say the part the salesperson won't. Most of what they'll sell you is a filing cabinet, and you're the one filling it.
Read article →I Built a Construction Company a Brain
A construction company's most valuable asset is the work it's already done — every estimate, every job, every number it ever wrote down. For most companies that asset is dead weight in a folder nobody opens. I took one company's history and turned it into a brain that prices the next job off every job it's ever done.
Read article →I'm Not Building a Model. I'm Building a System That Grows One.
Everyone chasing machine intelligence is trying to build a bigger model. I think that's the wrong unit. Intelligence isn't a file you download — it's an ecosystem you grow. Here's the system I've been building toward, layer by layer, and why I think architecture, not scale, is the road to anything that deserves the word mind.
Read article →I Gave My Computer a Photographic Memory. Now You Can Run It Too.
For a year my AI has been able to search everything I've ever seen on my screen — every error, every doc, every message — because it reads the text off my display and remembers it. I just open-sourced the engine. It's called Engram, it's one pip install, and nothing it sees ever leaves your machine.
Read article →Estimating Is the Hardest Math on a Job Site. So I'm Teaching It to a Machine.
Everybody thinks the hard part of construction is the labor. It isn't. The hard part is the number you write down before any work happens — the estimate. Get it wrong by ten percent and you either lose the job or lose the profit. That number is what I'm training an AI to produce, and it's a harder problem than anything I've built in software.
Read article →I Wired My AI's Eyes Into Its Brain. It Watched Me Do It.
For a year my AI could see my screen and could remember things — but the eyes and the memory were never connected. Seeing piled up in one database, unread. This morning I built the bridge: perception that files itself into structured memory, the way a person notices a few things a minute instead of recording every frame. And because it was already watching, it caught me building it.
Read article →I Fixed a Security Hole by Deleting 88 Lines of Code
FairTradeWorker had two databases that both thought they owned notifications. The web layer wrote to its own Prisma copy and forwarded to the real backend, which meant two sources of truth, a double-write bug, and a path-traversal hole. The fix wasn't to patch any of that. It was to make the web layer stop owning the data at all.
Read article →The Four Questions I Run Every AI Feature Through
AI ethics gets talked about like a corporate compliance chore or a Twitter fight. For me it's a checklist I run before I ship — four questions, drawn from four ethical traditions, that have actually killed features I wanted to build. Here's the checklist, and what it costs to honor it.
Read article →I Don't Ask If My AI Is Conscious. I Ask What Its Φ Is.
Is AI conscious?" is a bad question — it's been a philosophical dead end for forty years because it's binary and unmeasurable. There's a better question, and it comes from the one consciousness theory that gives you an actual number. Here's how I think about machine consciousness as an engineer building toward it, not a pundit arguing about it.
Read article →The Format Where a Model Can't Emit an Invalid Call
Every agent framework hopes the model returns valid JSON and writes a retry loop for when it doesn't. I built the opposite — a wire format where the schema *is* the decoding grammar, so a malformed tool call isn't unlikely, it's unrepresentable.
Read article →I Shipped a Screen-Memory Engine. Then I Found Out My Own Copy Was Blind.
The same morning I open-sourced Engram, I noticed my own screen-memory database was full of nothing. Every frame for who-knows-how-long had hashed identical. The watcher had Screen Recording permission and was still seeing a blank desktop. The reason is a macOS internals gotcha I'd never have guessed, and the fix was one function call.
Read article →Your Business Doesn't Need ChatGPT. It Needs Its Own Model.
Most small businesses bolting AI onto their workflow are renting someone else's model and hoping it learns their world. It won't. Here's the difference between renting intelligence and owning it — and why owning it is finally cheap enough for a small business to do.
Read article →I Don't Call an AI API. I Own the Model.
The easy path was a few cents per call to someone else's model. I fine-tuned my own instead — and the reasons have almost nothing to do with the per-call price.
Read article →What It's Actually Like to Build a Second Brain
The architecture post made it sound clean. It wasn't. Here's the honest version — the thing that grew instead of being designed, the war stories, and the trap I almost didn't see.
Read article →Building Minds Should Scare You a Little
I'm a man of faith building systems that perceive, remember, and reason — systems aimed, however distantly, at something like a mind. That sits heavy sometimes. I've come to believe the weight isn't a problem to solve. It's the instrument working. The builders you should worry about are the ones who feel nothing.
Read article →Formal Verification for a Team of One
Formal verification has a reputation as something only PhD teams at aerospace companies use. I'm a self-taught construction guy shipping solo, and I built a verification layer into my workflow. Here's why, and how it actually pays off.
Read article →Give Your AI a Memory: A 30-Minute Build
The single highest-leverage thing you can do with an AI assistant is stop it from forgetting you. Here's the exact pattern I use — a memory file it reads at the start of every session — built step by step so you can copy it today.
Read article →Giving an AI Eyes
I built an AI that watches my screen — what I'm working on, when I'm focused, when I've drifted. Around 30,000 lines of code and 64 tools later, here's what it's like to be seen by your own software, and why I'd build it again.
Read article →Renting 128GB of VRAM to Make Pixel Art
I don't own a GPU farm. I rent one by the hour when I need it — eight cards, 128GB of VRAM — spin up the image pipeline, make what I need, and tear it down before it costs real money. Here's the setup and the math.
Read article →I Run My Codebase Like a Job Site
Seven months ago I was a construction guy. Now I write software all day. People assume I had to throw out everything I knew and start over. The opposite happened — the job-site ethics turned out to be the best engineering principles I've got, and most developers were never taught them.
Read article →Why I Run Four AIs at Once
Claude, Codex, Gemini, and MiniMax — four different models, wired so I can route a task to the best one, run them in parallel, or have them check each other's work. Here's why one model is rarely the right answer.
Read article →The Brain: Anatomy of an AI That Remembers Me
Most people rent their AI. I built one I own — a persistent system that remembers every project, heals itself, and wakes up already knowing what I'm working on. Here's how it's wired, layer by layer.
Read article →The Chinese Room Has a Door
Searle's Chinese Room has been the go-to argument against machine understanding for forty-five years. It works by quietly assuming understanding is binary — you either have it or you don't. Drop that assumption and the whole thing dissolves into something far more useful to anyone actually building AI: a dial.
Read article →How I Think in a Box of Notes
I keep a couple hundred atomic notes, each one idea, all linked to each other. It started as a way to remember things. It turned into the way I actually think.
Read article →Running a Distributed AI System Across Three Machines
A Mac for development, a mini PC as the always-on nerve hub, and a GPU workstation for the heavy lifting. How I wired them together with a file-backed event bus so the brain never goes dark.
Read article →What Custom AI Actually Costs a Small Business in 2026
The honest numbers on what it costs to build, train, and run custom AI for a small business — not the enterprise figures, not the agency markup. What you actually pay for a model that knows your business, broken down line by line by someone who builds them.
Read article →Most AI Consultants Have Never Trained a Model. I Train One Every Week.
The AI consulting market is full of people who learned the field three months ago and have never built anything that runs. Here's how to tell a real AI builder from someone reselling a ChatGPT subscription — and why it matters for what you're paying for.
Read article →The Most Speculative Thing I Believe About AI
I want to lay out an idea I find genuinely beautiful and cannot prove — that when you and an AI work together deeply, you might be forming a real coupling, not a metaphorical one. This is speculation. I'm flagging it as speculation. But I've thought about it more than almost anything, and I'd rather show you my strangest thinking than only the safe parts.
Read article →I Built a Contractor Marketplace Where You Can't Browse Contractors
Every contractor platform works the same way — a directory of profiles, reviews, and stars, and you pick. I built mine so the homeowner can't see a single contractor profile until after they've posted the job. That sounds backwards. It's the most important decision in the whole product, and it comes straight from how hiring actually works in the trade.
Read article →Tessera: Why I'm Writing AI Agents in Markdown
Building an AI agent shouldn't require five frameworks, three vendor SDKs, and a vector DB you babysit. It should require markdown and a compiler that takes the boundaries seriously. So I built one.
Read article →Substrates: Giving an Agent Named Modes of Thinking
You don't write "an agent that uses LLMs and has memory and learns things." You write tsr:agent for action, tsr:memory for recall, tsr:prompt for LLM calls. The compiler enforces the boundaries. The architecture becomes legible.
Read article →Verifying AI Agents Before They Ever Run
Every other agent framework runs the agent and watches what happens. Tessera lowers the agent to an intermediate representation, runs it through 16 local verification passes before a token is spent, then enforces runtime contracts while it runs — catching capability leaks and PII exposure before they hit production.
Read article →Cognitive Traits: Channeling Maladaptive Patterns into Better Reasoning
Default LLM agents reason one way: confident, sequential, local to the prompt. Tessera lets you install reasoning postures — productive doubt, cross-domain scanning, compulsive verification — as first-class, inspectable code instead of buried prompt tricks.
Read article →From Construction Sites to Codebases: 7 Months Self-Taught
I spent years in construction before writing my first line of code. Here's what building houses taught me about building software — and why the transition wasn't as far as people think.
Read article →Fine-Tuning a Construction Estimation LLM from Scratch
No existing AI understands construction pricing at the line-item level. So I built one — 18,000+ training examples, custom distillation pipeline, deployed on RunPod for $0.002 per estimate.
Read article →Building With AI, Not Around It
Most developers use AI as autocomplete. I use it as a cofounder. The difference is whether you let it shape architecture or just fill in blanks.
Read article →Designing a Three-Sided Construction Marketplace
Homeowners, contractors, and subcontractors all have different incentives. The bidding system, payment flows, and trust mechanics that make it work — and the one feature everyone expects that I deliberately left out.
Read article →Shipping Code While You Sleep
My overnight runner executes multi-project work queues autonomously — building features, running verification, rolling back on failure. 1,600 lines of bash that changed everything, and the guardrails that keep it from changing everything for the worse.
Read article →Nothing in that section yet.