I Reverse Engineered Three 6-Figures AI Startup with NotebookLM and Claude
How I built a 3-agent NotebookLM system that turns Claude Code into a founder-matching engine, trained on 3 founders who hit six-figure ARR in under 12 months.
I was scrolling X at 2 AM. Coffee cold. Baby in the other room, just born.
My main client cut my hours again recently because of the economy.
A tweet stopped me. “Another solo founder just hit $500K ARR with Claude Code.”
I have Claude Code and even a Max plan.
I’ve built 40+ AI agents with it, and my ARR is nowhere near $500K. Not even close.
Can these things be formalized? So I started looking at NotebookLM as the missing layer.
So I did my research.
Anthropic crossed $7B in yearly revenue in 2025.
Claude Code itself hit $2.5B in nine months.
Founders using AI tools are reaching six-figure revenue 4.2x faster than founders who don’t.
The question kept me up. Where am I doing it wrong?
So I dug deeper.
I picked three founders who hit six-figure ARR with Claude Code in under twelve months.
Different niches. Different distribution. Same builder.
Then I built a system to interrogate them.
Who Are These 3 Founders?
Nomiki: Theanna ($207K ARR) Non-technical solo founder in Columbus, Ohio.
Builds a platform for women founders using Claude Code as her frontend engineer, Lovable for rapid prototyping, and Linear for handoff to one part-time engineer. Distribution is TikTok built-in-public plus SEO blog.
Shipped a mini-product called Marble Jar in 3 hours.
Jake Ward & Lara Acosta: Kleo + Mentions ($744K ARR)
Technical duo. Jake builds, Lara markets. Combined LinkedIn audience around 480K at launch. Forked the Vercel AI Chatbot template instead of building from scratch. Stack runs on Next.js, Vercel, Neon, Inngest, Clerk, Claude API, Claude Memory. Combined MRR hit $62K in three months.
Mickey: Late ($480K ARR)
Solo technical founder. Social media scheduling API. Built with Claude Code. One distribution channel only: Google Search Ads. Spends $8K a month, generates $15K MRR. CAC stays under 30% of first-year revenue. Hit $40K MRR in seven months.
Three different stacks. Three different distribution playbooks. One shared builder.
The Architecture: NotebookLM + Claude Code Agents
The architecture has 3 layers.
Layer 1 is the brain. Three NotebookLMs, one per founder. Each one trained on that founder’s blog posts, Indie Hackers writeups, podcast transcripts, and X threads. NotebookLM hallucinates less because it only answers from the sources you give it. If the answer isn’t in the founder’s own writing, the notebook says “Not in sources” and stops.
Layer 2 is the digital twins. Three Claude Code sub-agents. Theanna agent, Kleo agent, Late agent. Each one queries its own NotebookLM. Each one frames the answer in that founder’s actual decisions, with citations.
Layer 3 is the router. A local Python app on my laptop. It holds a summary map of all three founders: background, audience, capital required, and technical depth. I ask the router a question.
It picks the founder whose path matches mine. It routes the question to that agent. The answer comes back.
I picked three brains because one founder agrees with itself.
Three founders, three sets of trade-offs.
The router decides which trade-off matches the question.
What the NotebookLM + Claude Code Router Returns
First, the assistant will ask you a couple of questions.
And based on your answers, it’ll match you with one of these successful people.
Once you run Kleo, the agent will go to the notebook and start talking with it, like this:
This will prevent hallucination. And you’ll get the verdict, as you can see.
I asked too many questions, so there are too many verdicts. I couldn’t help myself :)
So, let’s build this system from scratch in four steps.
Step 1: NotebookLM (×3)
Three notebooks. Three brains. Each one was trained on one founder.