Building in Someone Else's Domain
In late spring 2025 I started working with an elite bodybuilding coach — champion powerlifter, hundreds of thousands of followers, twenty years coaching competitive athletes through contest prep. He was spending all his time coaching. Between managing his athletes, selling supplements, writing training programs, and answering the same questions over and over, he had no bandwidth left. He needed to scale what he knew without scaling himself.
So I started reading. Everything he had. Training protocols, nutrition frameworks, supplement guides, contest prep methodologies that he'd refined over two decades of working with the best. Dense PDFs, Word docs with no consistent formatting, years of FAQ exchanges with clients. Coaching gold mixed in with personal details that couldn't be exposed. I was building a brain from his life's work — something an AI could draw from to handle first-pass answers, draft content, power workflows through N8N. But the work of building that brain was the work of understanding his domain.
Not every document was worth keeping. Some material was outdated. Some contradicted other sources — his thinking on nutrition had evolved significantly over twenty years. Deciding what to include meant understanding his philosophy well enough to know which answers represented his current approach and which were artifacts. I built a pipeline that parsed his client exchanges into clean question-answer pairs while stripping personal information. Chunked it, embedded it, upserted it into a vector database. This was with Sonnet 4.0 — the extraction work alone would be meaningfully faster with current models.
By the time the knowledge base was solid, I understood his business. Where his value was concentrated. How his clients actually used his expertise. What questions came up constantly and which ones required his direct judgment.
About a month in, the coach asked to talk to the brain directly. It was answering questions the way he would — pulling from his actual protocols, drawing on the nuance of twenty years of competition coaching. He saw it immediately. His audience would pay for this. He coached the very best, and his knowledge at a $30/month subscription could reach the thousands of athletes who'd never afford one-on-one time with him. He asked me to build a product around it.
I shipped the full SaaS in about six weeks. Streaming chat interface with tool calling — nutrition plans rendering as tables, program recommendations referencing his specific protocols by name. Auth. Stripe. A separate coach-client portal for private messaging and resource management. The interface had to feel as polished as ChatGPT or Claude. That's the baseline expectation now — you can't ship a chatbot that feels like a side project.
The guardrail work took as long as the product development.
This is bodybuilding. The audience includes competitive athletes who use performance-enhancing drugs. They want specific, accurate information — not a disclaimer that says "consult your doctor." If the AI refuses to engage, the product is worthless to the people who'd actually pay for it. If it engages carelessly, someone gets hurt. The coach had twenty years of experience guiding athletes through exactly these topics. The question was how to deliver his actual expertise with appropriate context rather than suppressing it because it's sensitive.
Every guardrail was a product decision refined through dozens of iterations. Adversarial prompts, edge cases, conversations that started simple and escalated. Each round of testing surfaced gaps. We built a body composition calculator and program-specific tools so the agent felt congruent with the coach's existing courses. Which tools to build, when the agent should use them versus answering from the knowledge base — product design implemented as code.
Over a hundred paying subscribers in the first month.
The chatbot wasn't even the most immediate win. The knowledge base unlocked the coach's content strategy overnight. He had twenty years of material and a massive audience hungry for it — he just didn't have the hours to produce it. Once the brain existed, that bottleneck disappeared.
We parted ways after about three and a half months — the scope had outgrown the engagement.
If I did it again today, the code would take a third of the time. Six weeks of product development might compress to two or three. I'd still spend the same amount of time reading the material. I'd still spend the same amount of time on guardrails. I'd still start with the knowledge base.