A live, AI-powered SaaS product, designed and shipped solo.
CouncilAds is a working web app that shows advertisers which version of their ad lands best, and why, before they spend on clicks. A user submits their product. The app writes competing ad versions, and a council of twelve AI buyer personas scores and critiques each one. It then returns the stronger ad, a review of the landing page it points to, and a plain-English campaign setup guide.
It is live, it takes real payments, and one person built it end to end.
Most small advertisers write an ad, launch it, and find out it was wrong only after paying for the clicks. Testing a message with real money is slow and expensive, and it teaches you very little about why an ad failed.
CouncilAds moves that feedback earlier. Twelve modeled buyers react to your ad, each with their own priorities, doubts and triggers, so you find out what works and what doesn't while changes are still free.
The heart of it is the council. Twelve distinct AI personas, each modeled on a different type of buyer, read your ad and react honestly. One might say the offer is unclear. Another might say the headline promised something the ad never delivered. You get named reactions with reasons, not a vague score.
Around that engine sits a full product:
The first commit was June 12, 2026. Stripe paid plans went live by July 6, 24 days later. Since then the build has been steady: 398 commits and 34 merged pull requests.
I have run the Council more than 120 times on my own business and test ads. That is how most of these features were found and fixed.
Modeling how customers react with AI is a growing field, studied under names like synthetic audiences and generative agents. The strongest evidence so far is a 2024 study by researchers at Stanford, Northwestern and Google DeepMind. They interviewed 1,052 real Americans for two hours each and built an AI agent from each interview. The agents reproduced those people's own survey answers about 85 percent as consistently as the people did when retested two weeks later (Park et al., "Generative Agent Simulations of 1,000 People," 2024).
I want to be precise about how that relates to CouncilAds. The study built agents from long interviews with real individuals. The CouncilAds personas are modeled buyer types, not interview-based copies of real people, so the 85 percent figure does not transfer to my scores. What the research does support is the broader idea: modeled audiences are good at directional signal, such as which message lands, what objections come up and which version is weakest. That is how CouncilAds is meant to be used, as a fast, low-cost read before you spend, not a prediction of results. Scores are deliberately tough, and a strong ad lands in the high 5s and low 6s.
This is not a mockup. Every piece runs against real production services.
Payments run through Stripe and were validated end to end with real transactions, not test mode. Sign-in uses real Google authentication. After a tester told me the standard flow looked untrustworthy, I moved it onto a dedicated domain (auth.councilads.com), which fixed a real barrier to signups.
There was also an invisible problem. The site began as a JavaScript app that search engines could not read, so it was effectively invisible to Google. I rebuilt how pages are served so search engines see full content, added automatic sitemap generation, and set the site to notify search engines the moment new content publishes. Indexing was confirmed on the live site, not assumed from the code.
I use my own tool on my own pages, and it does not always flatter me. When I ran my own landing page through the Council, I made what looked like two sensible fixes. Both made the score worse, from 4.9 to 4.8 and then to 4.5. The personas wanted real pricing for the next step and real credibility, not softer reassurance wording. Adding honest pricing and a rewritten bio brought the score up to 5.2.
The lesson is built into the product. A fix that sounds good is not proof that it works, so CouncilAds re-scores revised ads and labels small changes as normal variation instead of calling them wins.
I designed the product, made the architecture decisions, and directed the build in plain language, with AI writing and iterating the code under my direction. I use the same approach on every WorkLeapAI project.
The rule throughout was that nothing counts as done until it is proven live. Payments were tested with real money and refunded. The sign-in flow was checked against Google's real servers. Each release was reloaded and re-tested on the live site before the next step. That habit of verifying instead of assuming is why the product holds up.
CouncilAds is proof of what I can build. If I can design and ship a live product with payments, secure sign-in, an AI decision engine and real search visibility on my own, the AI tool or automation your business needs is within reach. That might be an assistant that answers customers around the clock, a system that takes repetitive work off your plate, or a custom tool built around how your business runs.
The method is the same every time: build the real thing, test it against reality, and put it to work.
Want to see what that could look like for your business? Take the free 3-minute assessment, or get in touch through the contact page.