One Person. One Year. 12,891 Commits.
Here is the literal receipt: four git repositories, one author on nearly every commit, 12,891 commits across eleven months. No co-founders, no engineering team, no contractors. Just one person and a team of AI agents.
I started this company with a simple observation: most data analysis that small businesses need is the same dozen statistical methods applied to slightly different datasets. K-means on your customer segments. Linear regression on your sales drivers. Cohort retention analysis on your signups. The insights are highly specific to your business, but the analytical process is almost entirely repeatable.
If the process is repeatable, it can be automated. Not approximated with a dashboard — actually automated end to end. Upload your data, describe what you want to know, get a real statistical analysis with charts, confidence intervals, and a plain-English explanation. That's the product I set out to build.
The First Commit
The very first commit in this codebase is dated June 17, 2025 in what is now the api-mcp-prod repository. It says "update." Not an auspicious start. But the idea underneath it was clear: build the server as an MCP tool, so any AI assistant could call it. Instead of building a standalone app that users navigate, make the analytics available as a capability that plugs into whatever the user is already working in.
Six months later, the platform had a working pipeline, a production API, a webapp, and a marketing site. All four repos were active. The modules were being built faster than I could track them manually.
What the Pipeline Actually Means
The first analysis hit the registry on April 12, 2026. Within weeks the pipeline was building new analyses faster than I could track them by hand. Each one is a real analytical question a real business might ask: employee attrition drivers, fraud detection patterns, customer segmentation, geological factor importance, exam score clustering.
I am not building these modules by hand. An automated pipeline does it. When a new dataset arrives, a series of AI agents — each specialized for a different stage of analysis — plans the approach, writes the code, executes it, verifies the output, builds the visual report, and deploys it to production. My job is to specify what "a good analysis" means and to fix whatever the agents get wrong.
The pipeline runs on a Mac mini sitting on my desk. It costs about $15/week in Anthropic API credits to operate. It files a morning status report every day before I'm awake.
The Operating Model
Here is something I didn't expect when I started: the hardest part of building with AI agents isn't writing the prompts. It's designing the coordination layer. Agents need to hand work to each other, escalate when they're stuck, and accumulate knowledge over time so they don't solve the same problem twice.
The current system runs seven domain agents on the mini alongside the pipeline daemon:
The agents collectively produce something that surprised me: they get smarter over time. Working memory accumulates. Patterns identified in one session get propagated to the next. When mini-it spots a new error pattern in the pipeline logs, it files a TODO, which the pipeline-patcher agent picks up and turns into a code fix, which goes through review before I merge it. The loop closes without me in every step.
Why One Person
There is a standard story people tell about AI replacing workers. I don't find that story interesting or accurate. What I find interesting is this: agents let one person operate at a scope that previously required a team, without sacrificing quality or institutional knowledge. The agents don't replace anything — they multiply what one person can track, maintain, and improve simultaneously.
I can work on the module pipeline in the morning, knowing the agents filed the overnight error report, triaged the failing modules, and proposed the two code fixes that most likely to clear the queue. I don't need to reconstruct context when I sit down. It's already there.
That's what 12,891 commits and one author actually represent. Not heroics — leverage.
The Receipt You Can Use
Everything I've described is operational. The operating model is being extracted as embeddedlayers/agent-os — an open-source framework for running a company on AI agents. The full release ships in June 2026. It includes the coordination layer, the agent specifications, the memory format, the typed handoff protocol, and todo.sh, the 3,000-line Bash script that runs work coordination across the whole system.
The analytics product itself is the concrete proof this model works. MCP Analytics doesn't just tell you that AI-native analytics is possible — it runs on the same architecture it's selling. If you want to see what a real statistical analysis looks like, the sample reports are live. If you want that analysis on your own data, that's the product.
The next Build Log entry covers the module pipeline in more detail: how each agent stage works, what the common failure modes are, and what a good analysis actually costs to produce. It goes up next week.