About Bill's Web Application Demo Site

We’re all aware of the hype and drama surrounding AI. There are real issues related to privacy, energy/water consumption, bias, job displacement, IP, misinformation - all of which needs to be properly addressed before some organizations can comfortably run with AI. Given the complexity and gravity of those issues, I felt it appropriate to withhold my endorsement of AI until the point at which I could actually see first hand the benefits of an AI coding “partner” in custom development. Here's hoping we resolve those issues quickly as there is much to gain. In reality, it has taken me longer to document my thoughts on this page than it took to create a distributed, fully functioning, data-driven, professionally hosted Restaurant Menu Profitability Report displaying a quadrant scatter plot a restaurant operator could use to identify a sub-par menu item. All told, AI coding helped me create six sample dashboards and a demo site in less than one week. I also feel confident that as engineers, we can more effectively enforce coding, architectural, documentation and tool standards across development teams by standardizing prompting.

That said, I do not mind sharing my approach and a few observations I made along the way.

1. For each industry area, I asked an AI coding assistant to act as an industry consultant/expert to generate three best-practice management report suggestions often utilized to improve quality and/or profitability by monitoring key operating metrics. I provided some domain-specific examples ("multi-shot" prompting) of metrics and reports I considered "best practices" but I also encouraged my AI coding assistant to go beyond my suggestions.
2. I shared with my coding assistant my preferred programming languages and asked it to suggest popular tools and frameworks often used for data-driven web-based management dashboards. My ask was for AI to first write prompts, one for each suggested management report, and asked it to research randomized, publicly-available datasets I could use for each such dashboard.
3. Of course I asked AI to provide me clarifying questions before proceeding. I also emphasized the need for clear code comments, comprehensive testing and detailed “how to” “README” files where design decisions were captured in great detail.
4. I asked it to weigh in on which reports and associated metrics/variances, etc. were most feasible given likely constraints and limitations in available data. I was impressed with AI’s feedback identifying dataset gaps and data quality issues likely to impact the accuracy of certain reports.
5. The March Madness Prediction Model app in particular requires multiple data downloads, imports and prediction runs both before and after the actual NCAA tournament each year. To account for data and process complexity, I partnered with my coding assistant to create clear SOP's and validation processes to ensure the correct dataset is imported at the right time.
6. I then asked it to create prompts that outlined project implementation as well as describe high-level application architectural/design decisions (i.e., backend, data layer, frontend rendering, hosting, etc.). I emphasized that no code was to be written before I approved the design.
7. After iteratively reviewing and improving the AI prompts ("multi-turn conversations"), I finally began submitting these prompts to AI in a specific order. This approach gave me the opportunity to make design changes and improve the prompts along the way. I will say that the coding assistant did forget to apply some of the design ‘tweaks’ I implemented to reports implemented later in the process. I think I needed to embed those tweaks in the prompts themselves.
8. The AI coding assistant defended design/architecture recommendations and even presented a case for feature change based a dataset limitation discovered after implementation had started. The project structure, code comments, function sizes, naming conventions, stylistic decisions, and design documentation all look appropriate.
9. Extensive tests were automatically run each time an application changed and before code was committed to GitHub. As of October of 2026, more that 600 test scripts were authored by AI. That said, like us mere mortals, the AI code assistant will make mistakes. Examples include nul column values resulting from erroneous data import logic, improperly formatted styles and HTML code, some wonky layouts and sizing of images, and other mistakes commonly made by developers. Often AI discovered and fixed errors on its own before I noticed, and, in other cases, identified root cause and solutions after I called an issue out.
10. One lesson learned from the March Madness prediction model is that ideally I should have manually calculated a prediction with error rates using a simple data set and limited historical data before coding against the real data source. Then I would have compared the manual results to the application result in hopes of identifying logic and data issues earlier in the process. In general, the more you ask your AI assistant to rerun and reanalyze it's output, the better off you will be.

If you wish to discuss further, I can be reached by email: (enable JavaScript to view)