All posts

How Claude helped me build ACM.ski on AWS

In short

Claude Code wrote the code, and I led. It all runs on AWS, serverless, and the result is ACM.ski, for following alpine ski athletes from one season to the next.

Night desk: small robot with a wrench, cloud architecture and dashboard screens, lit ski slope outside the window

“My kid improved: he was 3 seconds behind the winner, and now he’s only 2 seconds behind!”

I’ve often heard that. The problem is that a second isn’t worth the same from one race to another.

I wanted to follow every athlete’s performance from season to season. With a real measure.

ACM, short for Athletes Career Management, was born from my two passions: skiing and tech.

💡 The idea

A CRM lets you track every client. I wanted the same thing for athletes: a profile, a history, a progression.

To compare different races, each result becomes a percentage of the winner’s time. That’s the ACM score.

An athlete often starts the season weaker than they finish it. This tracking lets the coach adjust things.

🎯 Starting from the end

I followed Amazon’s Working Backwards method: start from the result you want, then work back.

On October 1, the plan was set: testers at the end of October, then a public launch at the end of November.

Having a date helps you choose what really matters.

🤖 Claude wrote the code, I led

Claude Code wrote almost all of ACM’s code.

For programming, I start with Claude Opus. Sonnet did its share too. For everyday questions, I use Haiku instead.

I described what I wanted. Claude proposed the change. I accepted it, or had it fixed.

I felt more like a project manager than a programmer.

Nothing shipped without my go.

☁️ A serverless AWS architecture

Everything runs on AWS, in my own account.

No servers to manage. Services start only when they’re needed.

There are three environments: development, acceptance and production. Every release goes through the first two before it reaches acm.ski.

Claude also described the whole infrastructure as code, so I can rebuild any environment exactly the same way.

📄 Reading the results PDFs

At first, classic tools could barely read 1 test PDF out of 22.

With Claude on AWS, we got to 22 out of 22.

But AI isn’t perfect. On one PDF where the first name came before the last name, it swapped them for 40 athletes out of 40.

So we wrote better rules. Today, they read 91% of Québec PDFs without AI, and Claude steps in for the hard cases.

And a person validates every race before it’s published.

🏁 The result

More than 6,600 athletes. More than 2,150 competitions. More than 72,000 timed results. 53 clubs, 9 regions, five seasons.

It’s my first real project that I’ve put into production.

And it’s only the beginning: AI is coming this winter.

🍳 The simple version

  1. I start from a need: compare athletes with a real measure.
  2. I start from the end, with Working Backwards.
  3. I lead, and Claude Code writes the code.
  4. Everything runs on AWS, serverless, with three environments.
  5. Rules and Claude read the PDFs, and a person validates every race.

Check it out at www.acm.ski.