How MyTrainer AI was built: From Zapier to Agentic AI

A Framer landing page and a Zapier workflow (late 2023)

Back in end of the year 2023, a friend told me to try Framer to build websites. The first idea that came in mind is to build a simple AI workout generator. ChatGPT had exactly one year — it was still very new at that time and there were no agentic application. However, its API was already opened and people were trying multiple applications above. I selected a website template in Framer and built https://www.train-programs.ai landing page. Then, I integrated a Typeform and created a workflow in Zapier. It was triggered by the Typeform completion webhook, generated a workout program using the OpenAI API (was probably GPT-3.5 at that time) and send it by email using the Gmail plugin. Fairly simple, but end-to-end working workflow. After validating my workflow, I made a post on Reddit and shared it with some friends. It was pretty cool to have a personalized routine by email in a few minutes and for free. But that was it.

People were using it daily

A few months later, I noticed something interesting. While I was working on different projects, there were people that were using this tool daily. Not monthly. Not weekly. Daily. And it was growing: by June 2024, 3 to 5 people completing the form until the end with email delivery. It generated some AI costs but it was minor. So I decide to have a look at it again. First, I added a payment of $5 at the end of the form to receive the program by email. It didn't work out, people were not paying. It wasn't different enough than ChatGPT. Here, you might think I stopped the project. Well, I didn't.

A friend at Meta: build a mobile app

I talked to a friend working at Meta. He's a great mind — very strong in multiple fields, including AI and product. He told me something straightforward but that I didn't considered back at the time: people at the gym use their phone to keep note of their workouts — so you better build a mobile app if you want to move the project forward.

Here we are. I hired a UI/UX designer and we made the first mockups of what would become MyTrainer app. It didn't have nutrition or chat with MyTrainer AI. It had a similar onboarding than the website workout generator, a similar backend workflow (although I built a custom backend — not the Zapier's one that was expensive at scale) and user friendly sessions and calendar. I redirected traffic from my website to the app and got my first users like this (the website was still bringing traffic).

Why I bet on generative AI

I've always been a strong AI believer. After my engineering school, my internship in 2019 was about NLP (Natural Language Processing): I trained an AI model to detect named entities in specific customer support transcriptions chat for a major telecom company to conform with the GDPR. It was 3 years before ChatGPT. When building the very first version of MyTrainer, I had the strong conviction that ChatGPT was only the first wave before a tsunami. I bet on the rise of generative AI: increased quality, speed & cost as well as integration.

The vision

The vision of MyTrainer is simple: an AI personal trainer that knowns everything about you, the latest science-backed recommendations and guides you through your workout, nutrition and life in order to optimize your health. It's much more than gym: it's about longevity.

Sure, the static form to email delivery GPT-3.5 generated workout was far from this goal. But I decided to adopt a step-by-step strategy and targets workout generation first to scale as it's something people actually use (a lot of apps are doing this). The more users I would get, the more resources I would have to push the vision forward.

Shipping step by step — and a first agentic architecture

Speaking about resources, I couldn't afford to work full time on MyTrainer at that time. I started working in freelance on multiple very interesting projects, mostly in AI, and continued to maintain MyTrainer. I added new features, the biggest one of them being the integrated chatbot. I even created my first agentic architecture: I found out that AI can return structured JSON outputs and built my own tools library by passing functions schema to the system prompt & making it return the function called with arguments, if any function should be called. Functions were: create_session(), modify_session(), remove_session(). OpenAI SDK tools did not exist yet. It was cool, but as models were not very smart compared to today and not trained and optimized for agentic applications, tools usage was not efficient.

Meeting Gauthier

I continued shipping and learning new stuff in parallel. I talked about my app. Then, in early 2025, I met a fitness coach, gym owner and influencer that shared the same vision about AI in fitness. We started to work together on a new app version with major improvements: improved AI workflow, new nutrition recommendations, better AI results. I'll explain in another article how Gauthier helped me for the LLM evaluations I ran.

We worked hard and then shipped the new version of MyTrainer. It's basically the current application without all the optimizations made (app on IOS and Android). Feedbacks were good — but the app's infrastructure had some stability issues and Gauthier couldn't promote it further as he was busy running a company he founded that owns multiple gyms. So again, I worked on other projects to make money and learn and maintained MyTrainer for the current and new users that were still enjoying the app.

March 2026: refocus, Codex, and 1,000+ monthly active users

Recently, in march 2026, after I took some time off for personal reasons, I decided to focus again on MyTrainer. With new development tools like Codex or Claude Code, I've made huge software improvements both in app frontend and AI backend. I then focused on distribution and the number of users quickly rise to more than 1000 monthly active users. Reviews are good and I contact every single user to understand what they like and what can be improved. I will also detail product optimizations and growth strategies I used to get users with low budget (no ads or social media for a B2C app) by ranking on ChatGPT.

Where MyTrainer is today

And here we are now — a fully AI agentic native application with multiple tools, complex orchestration layer, personalized notifications and much more.

This was the first technical article of a serie about how MyTrainer AI was built. I'll dig as much as I can into the technical details in the next articles. I hope you enjoyed it!

Guillaume