Form idea to production: SISU road map.

1. Initial idea and motivation.

The project was born from a personal need. As a frequent gym-goer, I noticed the lack of a reliable platform to log and track physical progress. In Colombia and Latin America, available options were very limited and usually priced in USD or EUR, which motivated me to build an accessible solution adapted to our local reality and priced in COP.

2. Research and design.

The research was mainly qualitative and team-based. I spoke with trainers and other gym members, gathering direct feedback. Through these conversations, I confirmed the problem was widespread. I initially discussed the idea with ChatGPT, explained the problem, and it suggested which tools I could use, recommending Node.js and Express for the backend and React for the frontend, since at that time I only knew PHP and MySQL. I also explored NoSQL databases like MongoDB to optimize data storage, again based on ChatGPT’s recommendations.

3. Initial development.

Although my previous experience was in PHP, I decided to adopt JavaScript for development, guided by AI. With the rapid progress of AI, I initially wanted to build it in Python, thinking it would be better for handling large amounts of data and calculations. I tried for about a month, but ran into many issues and bugs. Since I was already programming using Cursor, it suggested that managing the project in JavaScript with Node would cause far fewer problems. So I switched to JavaScript, which significantly eased integration and enabled faster, more agile development.

I relied on multiple AI models. I would explain the problem to different models and ask them to generate highly specialized prompts, which I then passed to Cursor. Cursor allowed me to choose between models from Google, Anthropic, OpenAI, xAI, and others. I didn’t stick to just one model or use auto mode. If one model gave a response, I would delete it, try another, compare answers, and keep the one that was clearer, more consistent, and better explained. Over time, you naturally find a model that helps you the most—but you should never depend on just one. Even a preferred model can miss things that others catch.

4. Testing and validation.

The testing phase followed a trial-and-error approach. I used multiple AI models such as ChatGPT, Google, Anthropic, and others to debug issues and optimize the code. I would ask the same question or present the same problem to different models and keep the best answer. This comparison helped me find more effective solutions. Collaborating with AI was key to continuous learning and improvement.

5. Launch and continuous improvement.

The initial launch was limited to a close circle of users—family and friends—who provided valuable feedback. Based on their input, I refined and expanded features, adding elements like guides and advanced training options. I also implemented a feedback button to receive direct suggestions, ensuring the platform continues to evolve.

6. Lessons learned (the most important part).

A key lesson was not to rely on a single AI model; diversity of sources and constant research are essential.

GitHub was fundamental—vibe coding is trial and error, and GitHub is your safety net when things go wrong.

Always ask the model to explain what is being implemented and why; this leads to deeper understanding and better learning.

Test everything. Even if you have a favorite AI editor, the best answer often comes from the one you least expect.

Don’t get frustrated by bugs. Study them. Don’t just ask the AI to fix them—ask for a step-by-step debugging process.