Last July I hit my limit on Cursor. Literally. The free tier ran out and it told me I couldn't use it anymore for a month.
So I switched to Copilot. Except I didn't know how it worked. I didn't know I needed to pay $10 a month just to use it without also paying for the models underneath it. I didn't know what models even were, not really.
I was just a guy in Eastern Kentucky trying to build an app. No CS degree. No team. No funding. No idea what I was doing.
This was before Claude Code went mainstream. Before every AI company had a coding agent. Before any of this was normal. I was copying code out of ChatGPT, pasting it into VS Code, asking Claude questions in another tab, and manually trying to stitch it together. It was slow and painful and I kept hitting walls I didn't know how to climb.
So I decided to build my own coding assistant. Something that would actually help me.
Then I noticed something that bothered me.
The model I was paying for -- the expensive one -- was spending most of its time fixing ESLint errors. Formatting issues. Cheap problems. I was burning real money watching a frontier model argue with a semicolon.
I thought: why can't a cheap model handle the dumb stuff while the expensive model handles the hard stuff?
I asked the AI chatbots about it. They told me multi-agent orchestration was really hard. Agents working in parallel created all kinds of race conditions and conflicts. It would take a serious engineering team. Probably two years of work.
I decided to try it anyway.
I called it Maestro. Because it was conducting multiple agents like an orchestra.
Then I found out all the models cost money. Every single one.
So I thought: fine. I'll just make my own.
I started learning about distillation -- the process of training a small cheap model to do a specific thing well by learning from a bigger model. And when I heard the word "distill" the first thing I thought of was a still. Moonshine. The thing people in these hills made when the gatekeepers told them they couldn't have the real thing.
That's when I named it Moonshiner.
But I needed someone to check the quality of what came out of the still. You can't just drink whatever drips out. You need someone with a good palate who knows the difference between something worth keeping and something that'll make you go blind.
I named that quality checker Pappy. After the most respected name in Kentucky bourbon.
Pappy doesn't just say yes or no. He says PASS, WARN, or FAIL. With a confidence score. And if something fails, the loop runs again until it passes.
I kept building. I made a desktop app. I wrestled with MCP integrations until I got frustrated enough to build my own protocol for agent-to-agent communication. I called it the Agent Handoff Protocol. The runtime is called Mailman because it delivers the packets.
I named the access control layer Miranda. Because she reads agents their rights -- you can only use the tools you're authorized to use, nothing more.
I named the task router Brain.
I built a conversation layer and named her Benson.
I named the whole thing Orca. Because orcas hunt in coordinated packs, each one with a role, none of them wasted.
My wife named the engine and designed the logo. That's the whole team.
I shipped v1.0.0 in February 2026. Six months after I hit my limit on Cursor.
And here's the part that's been messing with my head ever since.
In the six months since I committed Moonshiner v0.5.0 on November 28, 2025, 31 external teams have independently shipped pieces of this architecture. IBM. Microsoft. AWS. Google. Anthropic. Alibaba. Meituan. Sakana AI. UC Santa Barbara. DARPA.
And yesterday -- Cursor. The exact tool that hit my limit in July 2025 and sent me down this road in the first place. They just shipped /orchestrate with planners, workers, and verifiers. If verification fails, the planner spawns a new worker to fix it.
That's Brain plus Pappy plus the repair loop. Drawn as a diagram. Posted by the tool whose free tier I couldn't afford.
Every single one of them built a piece of what I already had working.
Every single one of them was missing the same piece: the contractual relationship between Pappy and Moonshiner. The quality gate that feeds the distillation loop. The system that doesn't just fix the current output -- it trains a better model for next time.
I documented all 31 in a file called PRIOR_ART.md in the repo. Organized by component. With a gap matrix. With timestamps.
This wasn't the first time.
In June 2025 I had an idea for an app that used OCR to scan receipts and automatically track your warranties by make and model. Three different AI chatbots told me I needed to patent it. I didn't move fast enough. Six months later warranty tracking with receipt scanning was everywhere.
I didn't learn my lesson fast enough. But I did learn it.
When I started building Maestro I stopped asking for permission.
Some days I was sure I was wasting my time. There were no peers to ask. No investors to signal that someone else believed in it. No community of people building the same thing. Just me, a laptop, Eastern Kentucky, and a stubborn belief that the tools people were settling for weren't good enough.
The chatbots told me it would take two years. I did it in three months.
I was told multi-agent orchestration was too complex for a non-coder to build. I shipped 620 passing tests across 12 packages.
I was told you need money for models. I built a distillery.
Now Anthropic ships Outcomes -- a quality gate for agents -- and calls it a breakthrough. I built Pappy in November 2025. I built the part that comes after it too.
I'm not telling this story to be bitter. I'm telling it because somewhere right now there's someone being told their idea is too hard. It'll take too long. They don't have the skills. They should wait until they know more.
Don't wait.
The tools exist to build things that didn't exist last year. The walls people told you about are shorter than they look. The two-year timeline they quoted you is a guess from someone who never tried.
I'm one person. Eastern Kentucky. No funding. No team. No CS degree.
I built the thing 31 teams independently decided needed to exist.
You can build yours too.
There's one more thing I want to say.
Eastern Kentucky gave the world coal that powered the industrial revolution. Bourbon that became a global industry. Music that influenced every genre that followed. This region has always built things the world needed. It just never got credit and never captured the value.
I want to change that.
The long term vision behind all of this isn't just Orca. It's Silicon Holler -- an innovation ecosystem on the eastern side of this country where people who can't afford to move to San Francisco don't have to. Where the kid in a small Appalachian town who's too smart for the options in front of them gets a shot at building the future instead of just watching it happen somewhere else.
Every great tech hub started with one person who proved it was possible.
I'm not saying I'm that person. I'm saying November 28, 2025 is the day we started finding out.
The tools exist. The talent exists. The only thing missing is the belief that it can happen here.
I believe it can.
Repo: https://github.com/junkyard22/Orca
Prior art breakdown: https://github.com/junkyard22/Orca/blob/main/PRIOR_ART.md
AHP runtime: https://www.npmjs.com/package/@marsulta/mailman
Top comments (0)