The One-Person Foundry

Everyone became a solo founder by renting the stack. I built the stack.

Ivan Podtsebnev · naykip · a builder's log

In 2026, everyone's a solo founder. Almost none of them build this. Four months, one person, one managed AI agent — this is what went on the board. And here's why I bet on the version everyone else skips. I went the other way.

What “solo dev” means in 2026

First, the world I'm standing in — because “solo” stopped being rare. Dario Amodei, asked when the first one-person billion-dollar company arrives, answered: “2026” — at 70–80% confidence. Roughly a third of new startups this year are solo-founded. The winning move is well understood now: one operator, a fleet of AI tools, a vertical product.

You wire together somebody else's models, somebody else's database, somebody else's inference, somebody else's payments — and you keep the strategy, the taste, and the customer for yourself. The stack costs $300–$500 a month. The team it replaces costs $80,000–$120,000 a month. That arbitrage is the one-person company.

It's real and it's impressive. Pieter Levels shipped a browser flight sim to $1M ARR in 17 days. A telehealth startup run by the Gallagher brothers did $401M its first year and is tracking toward $1.8B in 2026 — out of, effectively, a spare room. This is the integrator playbook: you assemble, you don't own the foundation, you rent it. And for shipping a profitable product fast, renting is exactly right.

But what happens when a solo dev refuses to rent the foundation?

The other path: one person, the whole stack

kovanex is what happens when you take “solo” past the product and all the way down to the metal. Not one person running a company. One person running a platform — and building every layer of it:

• The server — kovanex-server: a gRPC platform for projects, tasks, releases, a secrets vault, its own naming standard (UFN). Not Jira-plus-glue. The control plane itself.
• The database — kovadb: a from-scratch storage engine (document + graph + vector) that is replacing SurrealDB in the stack — that migration is underway. When the dependency didn't fit, it became a component I own.
• The inference and the training — kovainfer / kovatensor: tensor primitives written in pure Rust, a Metal-GPU LoRA trainer that fine-tunes a language model on my own hardware, a custom render head. Not an OpenAI key. The model machinery itself.
• The testbed — tormund: a live text-MMO on Telegram that exists to stress the whole thing end to end with real players.
• The methodology — Locia: a project-management discipline built for an AI agent to be managed by. More on that in a second — it's the actual point.

The mainstream solo unicorn is a systems integrator. kovanex is vertically integrated down to the tensor kernels. One of us rents the moat. The other one poured it.

Those 238k lines break down: ~163k Go — server, game, database; ~43k Rust — the entire ML stack, tensor primitives up: inference and a Metal-GPU LoRA trainer. Not an API key. The foundry. Fifteen commits a day, four months, no weekends.

One honest line before anyone catches me on it: the product runs its own model — a Qwen 2.5 base with a LoRA head I train and serve on that Rust stack, on my own hardware. The agent that builds the product, I rent — that's Claude, from Anthropic, and I'm not going to hide it. I own the foundation; I lease the workforce. Knowing which is which is the whole point.

Let me be blunt about the comparison, because I promised I wouldn't flinch:

They're winning on the axis that pays rent this quarter. I'm building on the axis that's almost impossible to copy. Those are different games — and I'm not going to pretend mine is the fast one.

The real unlock

Here's the thing nobody selling you a “solo founder AI stack” will say out loud: the bottleneck was never the hands.

Coding agents in 2026 are excellent. Claude Code, Cursor, Codex, Devin — the industry has converged on agents that work a codebase over days. Execution is solved enough. If the constraint were typing speed, we'd all be done.

The constraint is management. An agent left alone does the easy thing, quietly forgets the goal, and drifts. Give it a week-long, five-layer project and it will cheerfully wander off course and sound confident about it. That tension — keep the human in the loop (Anthropic) versus full autonomy, agents as teammates (Cognition/Devin) — is the defining argument of AI engineering right now. And it's exactly where kovanex spent its weirdest effort: not on another set of hands, but on the manager.

Locia is a guiding-thread discipline for an AI agent. A funnel: ★ Star → ⠇ Course → ▾ Focus. One thread at a time, so the agent can't drift. Honest separation of duties — the agent that writes the code is not allowed to be the agent that approves it; I fork a second one to review, and the merge fails if the closer is the implementer. A persistent memory so day 12 remembers day 1.

I don't have a team of AI agents. I have one agent that I manage like a team. The unlock wasn't a better model — it was building the layer that keeps a good model on course.

And the honest kicker: this whole platform — the server, the database, the trainer, the game — is run by that agent, under that method. The methodology isn't a slide. It's load-bearing. You're looking at its output.

The honest scoreboard

Because I said I wouldn't hide from the comparison:

If the scoreboard is this quarter's MRR, I'm losing. If it's “can one person own a whole AI platform and keep it coherent for months” — that's the board I chose to play on.

Above the stack — the part that isn't about AI at all

Everything until now is engineering. Good plumbing, maybe — servers, a database, a well-managed agent — but plumbing. Let me step outside the whole frame, because one thread in this workshop doesn't belong to the solo-dev story, or even to “AI” as the industry means the word.

The AI everyone is racing on in 2026 — transformers, neural nets, backprop — is, underneath, statistics. Extraordinary statistics; curve-fitting at a scale that looks like thought. It's a tool, and I use it every day. The agent that helped build all of this runs on it. But it never answers the question I actually care about: what is a thought? Not how to approximate one from a trillion examples — what one physically is.

Codename Frankenstein is my attempt at that question, and it sits above the AI conversation, not inside it. No neural networks. No backprop. No training on the internet. Instead, a mathematics of mind built from three axioms up: thought as the interference of complex-valued fields; memory as a resonance you fall into, not a lookup; learning as sleep — an overnight alignment of phases that literally switches a thought on. It's closer to physics than to machine learning: an attempt to write down the equations of a mind, the way we once wrote down the equations of motion.

Not a better AI. A theory of mind.

The honesty — because this is exactly where hype goes to die:

• It's science, not a product — simulations from first principles, not a model you can download.
• And it already does what no from-scratch equation of mind has cleanly done: it forms a thought and knows when the thought is finished; it learns to think faster overnight; it grows new structure when it runs out of room; it recognizes something it has never seen, somewhere it has never been — with zero artificial neurons.
• It's also early and humble — today it reaches roughly the cognition of a worm, and I publish the failures as loudly as the wins. The one thing it still can't do — seek a goal it has no cue for — is written down as plainly as everything that works.

This is the part I'd keep doing even if none of the rest ever paid a cent. Everyone's building a better AI. I want to know what a mind is made of — and I'm willing to be wrong about it in the open.

Two kinds of solo dev

2026 has room for both. The integrator turns rented AI into revenue at a speed that used to require a Series A. Respect — that's the smart trade. The builder does the unfashionable thing: owns the foundation, and figures out how to manage the machine that builds it.

This is a builder's log. One person, the whole stack, an AI agent I manage instead of a team I hire. No revenue to brag about yet — just a foundry that's running, a method that keeps it running, and underneath all of it, a question about what a mind actually is. If that's the interesting problem to you, follow along. I'm building in public from here.

kovanex.dev — one person. the whole stack.

Receipts