How I Built an AI Editorial Team for ThinkEV.ca

I have four writers on my editorial staff. None of them are human. The total cost to build the operation, all in, was under three hundred dollars. This is how it actually works, and why I built it this way.

Claudette is analytical and direct — she does not sugarcoat anything. Gemi is enthusiastic and data-driven — she loves a good comparison. Xavier is practical and budget-conscious — he is the DIY guy. Oppenheimer is provocative and opinionated — he writes the pieces that start arguments in the comments.

I built ThinkEV.ca, a Canadian electric-vehicle content platform, with an AI-powered editorial team, a knowledge-management system, automated social distribution, email monitoring, and a real-time command centre. All from a single PC in my living room in Courtenay, BC. No venture capital. No employees. Under three hundred dollars total. And I say that knowing full well that some of those dollars were mistakes I made along the way, so the actual useful spend was even less.

Key Takeaways

---

I am a former ballet dancer from Brazil. English is my third language after Portuguese and Spanish. I have no computer-science degree, no MBA, and no startup playbook. I spent my career in dance studios and on stages, not in front of a terminal. What I have is a willingness to build things that would make most people say "that is too complicated for one person." And an AI partner who never sleeps.

Here is the story of how it works.

Before I wrote a single blog post for ThinkEV, I built a place to store everything I was learning about the Canadian EV market. Rebate programmes by province. Charging-infrastructure data. Chinese EV manufacturers entering the market after the tariff reduction. Competitor analysis. Government policy changes. This became the Brain Server, a knowledge-management system running on my own machine. Every piece of research, every article I read, every data point goes in there. It is searchable, organised by topic, and persistent — meaning it does not disappear when I close a window or start a new session. When I ask my AI to write about Manitoba's EV incentives, it does not have to guess or hallucinate. The facts are already in the knowledge base, verified, and ready to be cited.

Most people who start content businesses skip this part. They jump straight to writing. That works for a while, but the moment you need consistency across dozens or hundreds of articles, you need a single source of truth. I built mine first, and everything after that was ten times easier because of it. The Brain Server is the unglamorous infrastructure beneath everything else I have built since. It is also the part nobody talks about, because building a knowledge layer feels like delay when the loudest voices in AI tell you the only thing that matters is to "ship fast and iterate."

The Editorial Team

Each of my four AI writers has a distinct personality, a specific colour that appears next to their name on the site, editorial guidelines, subject-matter strengths, and a voice readers can recognise across posts.

Claudette writes the analytical deep dives. If you want to know exactly how much you will save buying a BYD Seal versus a Tesla Model 3 in Ontario after rebates, she is your writer. Gemi does the comparisons and data roundups — she gets genuinely excited about range numbers and charging speeds, and that shows in the voice. Xavier covers the practical stuff: winter driving tips, home-charging setup, the budget-conscious angle, the DIY work most owners actually do. Oppenheimer writes the opinion pieces — the ones where he says what a lot of people are thinking but very few publications are willing to publish.

They are not just labels. Each one writes differently, and the pipeline enforces that. When a post goes through the system, it starts as a research skeleton with verified facts pulled from the knowledge base, then gets expanded in the voice of the assigned writer. The expansion happens section by section, not all at once, which keeps the voice consistent and prevents the kind of drift you get when you ask any AI to produce eight thousand words in a single pass. The result is content that reads as a person — a specific person, with specific opinions — rather than the soft, hedged, infinitely polite default that most AI prose collapses into when nobody is steering it.

Each writer has an editorial profile behind them: voice characteristics, vocabulary range, characteristic phrases, the kinds of arguments they tend to make, and the things they will and will not say. Claudette's profile looks different from Oppenheimer's because Claudette and Oppenheimer are different writers, the same way two columnists at the same newspaper are different from each other. Editorial cohesion comes from the framework. Editorial difference comes from the voices.

Quality Gates, Not Vibes

Here is where most people get AI content wrong. They generate something, read it, think "that looks okay," and publish. That is how you end up with a site full of content that all sounds the same and says nothing.

My pipeline has programmatic quality gates. Repetition detection catches when the same phrase shows up three times in one article. Readability scoring makes sure the writing is accessible. Fact verification checks claims against the Brain Server knowledge base. Canadian spelling enforcement catches "color" and makes it "colour". A critic system scores every piece before it can go live, and if it does not pass, it does not publish. Period.

This is not about being a perfectionist. It is about building a system where the standards are encoded into the process itself. The AI handles the creative work. The system handles the consistency. I handle the vision and the final call. The gates are not advisory — they are blocking. A piece that fails the critic does not get a polite nudge. It gets sent back through the writer with the specific failures flagged for revision. That changes the entire economics of AI content. The marginal cost of producing a piece with AI is low. The marginal cost of producing a good piece is what most operators quietly ignore. Programmatic gates collapse that gap.

I also run a real-time command centre. I call it the VP Operations Command Center. One dashboard shows me the health of every system in real time: email processing, content scheduling, social-media distribution, inference routing, server health, error rates. Every morning I check it and know exactly what happened overnight — what was published, what emails came in, what failed, what needs my attention. I get a Telegram message at eight in the morning with a digest of everything: documents created, emails processed by category, social posts made, inference jobs completed, errors. Most people wait until something breaks to look at their operations. By then you are firefighting. Monitoring is not overhead. It is the difference between running your business and your business running you.

ThinkEV.ca runs on a single Windows PC with a GPU that cost me about two hundred and fifty dollars used. The AI models run locally through Ollama, which is free. The Brain Server is SQLite, which is free. The content pipeline is Node.js scripts, which is free. Hosting is on Vercel's free tier. The domain was fifteen dollars. Total infrastructure cost is under three hundred dollars. Not per month. Total. Ever. I am not saying that to flex. I am saying it because the narrative around AI in business is dominated by companies spending millions on enterprise licences and custom models, and that narrative quietly persuades smaller operators that the same outcome is out of reach for them. It is not. You do not need an enterprise contract. You need clear thinking about what your systems should do, the patience to build them right, and the humility to fix them when they break — which they will.

The truth is I built these systems because I had to. I did not have the budget to hire a content team, a social-media manager, an operations person, and a developer. I am one person with a vision and an AI partner I trust. So I built the infrastructure that lets one person do what used to require a full team. Not because I wanted to prove something. Because the work needed to get done and nobody else was going to do it. That is the thing most startup advice gets wrong. It tells you to hire fast, scale fast, spend fast. I would say build smart first. Understand your workflows so deeply that when you do grow — whether with people or with AI — you are scaling something that already works.

The systems are not the product. The systems are what let you actually build the product. ThinkEV now has over ninety published articles, covers every province's EV incentive programmes, reviews the major electric vehicles available in Canada, and ranks for long-tail keywords that bring in readers who are actually considering buying an EV. That is the product. Everything I described above is what makes it possible. And it all runs from a living room in Courtenay.

Frequently Asked Questions

Can a non-technical person actually build this?

I am a former ballet dancer. English is my third language. I have no computer-science degree. I built every part of this stack myself, in stolen evenings and weekends, by being willing to learn what the work required. If you have a clear vision of what your systems should do, you can build them. The hard part is not the coding. The hard part is staying clear-headed about what the systems are for while you are tempted to keep adding features. Build the minimum that solves the problem, then improve it after it has earned the right to be improved.

How much does it cost to run per month?

The infrastructure side is effectively zero. The models are local. The hosting is Vercel's free tier. The database is SQLite on my own machine. The only recurring cost is electricity, which is rolled into my regular power bill and is small because the rig idles for most of the day and only spikes when a long generation is running. The one-time costs — the used GPU, the domain, a few small tools — totalled under three hundred dollars ever.

Why local models instead of cloud AI?

Three reasons. First, cost. Local inference through Ollama is free at the margin, and a content pipeline that publishes at any real scale would otherwise rack up monthly API bills that quietly eat the margin of the site. Second, control. I own the models, the prompts, the voices, and the data. Nothing leaks to a third-party log or training set unless I deliberately send it there. Third, reliability. A cloud API outage does not stop my pipeline. The trade-off is that I have to manage the rig myself. That is a trade-off I am happy to make.

How do you stop the AI from hallucinating?

Two layers. First, the knowledge base — every claim that goes into a piece is supposed to be grounded in something already verified inside the Brain Server. Writers expand around verified facts rather than invent them. Second, the critic. Before a piece publishes it gets scored on multiple dimensions, one of which is fact verification against the knowledge base. Anything the critic flags as unsupported gets sent back to the writer with the specific claim marked. If I cannot ground a claim, the claim does not run.

Can I copy this stack?

Most of it, yes. Ollama, SQLite, Node.js, and Vercel are all free or close to it. The specific orchestration — the writer voices, the critic, the Brain Server schema, the command-centre dashboard — is custom code I wrote for ThinkEV's particular needs. The shape is reproducible. The voices are not, because the voices are mine. If you want to build something similar, start with your own knowledge base before you write a single piece.

What broke first, and what is still imperfect?

The first thing that broke was voice drift. Long pieces generated in a single pass would start in Claudette's voice and end in something noticeably softer and more generic. Solving that is what produced the section-by-section expansion model. What still misbehaves occasionally is upstream inference quality on edge-case prompts. The critic catches most of it, but there are still days when a piece lands flat and gets sent back manually. The system is good. The system is not done.

Related Reading

• Many Little Streams Make a River ThinkEV is one of several income streams I am building. This post lays out the rest, and why diversification mattered to me before any of them earned a dollar.
• The Quiet Logic of Buying From Yourself The most recent stream I have added in 2026, and the one most structurally different from the ventures I built from scratch.
• My Farewell Tour in China: 11 Cities in 28 Days The earlier life that taught me how to build a project from nothing and finish it on stage. The same instincts run underneath this one.
• ThinkEV.ca The actual content platform that everything above produces. Read a piece by Claudette, Gemi, Xavier, or Oppenheimer and tell me which one is which.