Behind the scenes

AI is fuelling our software development

Galaxus team Daedalus is moving more than a third of all data out of our monolith. With AI, this now happens so fast that an entire stage of the process has become superfluous. What makes this possible is preparation: the team laid the groundwork with agent skills and templates.

Why Exodus?

Much of what happens behind the scenes at Digitec Galaxus lives in a single database: Devinite. Customer orders, invoices and our purchase orders. For a long time, that had its advantages. One team could read another team's data directly and write to it in the same transaction. It was fast, simple and always consistent.

As we grow, this is starting to have a reverse effect. A large data change in the purchase orders locks tables other teams are waiting for. A schema change has to be planned with the Database Magicians, our database team, and sometimes carried out during a maintenance window in which our online shop is unavailable. And the other way round, every team feels the problems another team has with its tables. Nobody is truly isolated from anyone else.

Who's putting the load on Devinite? Share of database workload by team, measured over the seven days up to 9 November 2026. Team Daedalus leads with 13.6 per cent, almost twice as much as the next team.
Who's putting the load on Devinite? Share of database workload by team, measured over the seven days up to 9 November 2026. Team Daedalus leads with 13.6 per cent, almost twice as much as the next team.

That's why Team Daedalus launched what it calls Exodus: the purchase orders the team is responsible for will be moved out of Devinite and into a database of their own. The team can then evolve its data model and scale independently as well as archive old data whenever it suits them. Other teams no longer access the tables directly, but through clear interfaces they agree on with Team Daedalus. This makes Devinite smaller and easier to run.

In October, I worked with Team Daedalus as part of a DevXchange. In this internal exchange programme, you spend a few sprints in another team. In this post, I show how big Exodus is and how good groundwork and AI are moving the project forward so fast that a whole planned stage has become unnecessary.

How big is Exodus?

Devinite is the database of our monolith: 1,358 tables, around 21 billion records and 34 teams that own data in it. Our architecture vision is to split the monolith step by step into modules with their own databases. That way, every business area is clearly separated and controls its own data.

Who owns the 21 billion records in Devinite? Blue: purchase orders (Team Daedalus), 7.6 billion records. Grey: the 33 other teams
Who owns the 21 billion records in Devinite? Blue: purchase orders (Team Daedalus), 7.6 billion records. Grey: the 33 other teams

With the purchase orders, Team Daedalus is moving out a particularly large piece. The team owns 165 tables, more than any other team. They hold 7.6 billion records, more than a third of everything in Devinite, even though the respective code makes up just under a tenth of the monolith. The table with all the positions for products we've ordered from suppliers alone has more than 400 million rows.¹

And this data never stands still. While you're reading this paragraph, around ten new purchase orders will be created. On top of that, deliveries are booked in and invoices matched around the clock. So the rebuild happens while everything keeps running. Before the data can move to the new database, every access by other teams has to go through interfaces. At the start, code from other teams accessed Team Daedalus's tables directly in 47 places. Today, it's zero.

Daedalus' team space
Daedalus' team space

Five months of work in two weeks

Team Daedalus has eight members and works on several other large initiatives alongside Exodus. Even so, the team has pushed Exodus forwards with a lot of dedication. In the summer, it completed the first stage: splitting the creation and changing of purchase orders into separate transactions. Although there was considerably more work in this than expected, the team stayed within time and budget and finished the work in four sprints. Along the way, it classified every use case and made targeted use of AI agents to build the skills and tests that laid the groundwork for the big push from October onwards.

In October, Guillermo from Team Daedalus and I implemented further stories for Exodus, most of them as part of my DevXchange. That way, Exodus keeps moving while the team can focus on its business initiatives. In just over two weeks, we put 35 stories live – a good three per working day on average. Without AI, I would've estimated around three days for a comparable story, so around five months for all of them. Across Exodus as a whole, that means 148 of 211 stories are done, so around 70 per cent. Meanwhile, the skills and automation have come so far that the AI extends and refines the Exodus stories itself.

A well-coordinated team

Exodus isn't moving this fast because of AI alone. Team Daedalus knows the business processes around purchasing very well. It matters to the team that knowledge doesn't depend on a single person. That's why it takes a lot of time for refinements, in which it discusses new stories together. The team also follows fairly strict architecture patterns: the code mirrors the business language of purchasing, and reading and writing data are strictly separated – from the entry point through the business logic down to the database. What brings order for people also helps the AI. It finds its way around quickly because every part of the code is built to the same pattern.

A look at Team Daedalus' virtual Scrum board
A look at Team Daedalus' virtual Scrum board

This is Team Daedalus:

  • Noël, Team Lead Engineering
  • Emanuel, Senior Product Owner
  • Guillermo, Software Architect
  • Thomas, Senior Software Engineer
  • Samuel, Software Engineer
  • Gioele, Software Engineer
  • Nicolas, Software Engineer
  • Alain, Software Engineer

What the AI does

Much of this pace comes down to groundwork the team laid long before my DevXchange.

The foundations

Team Daedalus mainly works with Claude Code, an AI tool by Anthropic that reads and writes code and runs commands directly. Since the Opus 5.5 model, it's been good enough to take real work off the team's hands. Since May, the team has been using CLAUDE.md files. They explain to Claude how the architecture is structured and which rules apply. For example, to stick to few comments in the code, consistent naming, the team's architecture patterns and, when in doubt, checking how the team last implemented similar cases. Over time, the team developed further fixed workflows for Claude: research, plan, implement, review. Each step is a skill of its own, a set of instructions tailored to the team that Claude loads for exactly that step.

This helped with Exodus in particular. The team assigned every affected use case to one of three types: a switch changes the database connection, a split divides a write operation into two independent transactions, and an async makes a process asynchronous. For each of these types, Nicolas – my namesake in the team – wrote a skill of its own. With it, Claude implements the stories quickly and with fewer mistakes. On top of that, there are automated tests that check for every use case whether the change is correct. So the AI notices by itself when it has made a mistake and can fix it.

How AI implements stories

I tell Claude to read the story and its implementation plan. Claude finds the affected places in the code and proposes a plan that follows the team's conventions. I review and correct the plan and have Claude implement it. Then I look at the result, adjust things if necessary and open the pull request. Someone else from the team then has to review it. I mainly had to step in when a pull request got too big. Then we split it into smaller ones so it could still be reviewed properly.

This was most impressive when removing accesses from other teams' code. Claude tested the affected views in the ERP, our in-house management system, by itself: it looked up suitable purchase orders for testing in the database, started the ERP locally, compared the result in the browser with our test environment and fixed errors step by step. It was quite entertaining to watch Claude at work.

Where the AI gets it wrong

It's not flawless, though. Where the AI gets it wrong often shows up in the team's review comments. Once, Claude put logic into the purchase orders that actually belongs in another team's interface. Another time, Claude and I wanted to share code between two product areas. Guillermo, the team's architect, vetoed it. And another colleague found a real bug that Claude had missed. The AI can't replace the knowledge of how the team builds things and why. Sometimes it lacks an overview of the whole codebase and deep knowledge of every product area. Its context – the amount of information it can process at once – isn't (yet) big enough for that.

When work packages shrink

Exodus is well planned. The team split the project into stages, each as an initiative of its own – a larger project that a board including the executive management approves up front. The concept estimated four to six sprints per stage.

With AI, these estimates have shrunk dramatically. The first stage had classified every use case, and the skills and tests were ready. So the stories of the next stages could be implemented along the way. One stage has become unnecessary altogether: the stage for supplier invoices and cancellations, estimated at four sprints, was done before it had even started. In November, the team closed the initiative. Another one has largely been implemented while the team's still working on it.

That was only possible because the preparation was solid. AI didn't replace the concept, the classification of every use case or the team's refinements. They are what made the AI so effective in the first place. In the age of AI, preparation and refinement are therefore becoming more important, not less. They pay off many times over in implementation.

The new bottleneck

The bottleneck has shifted – to reviews. If two people together deliver a good three stories a day, someone has to read just as many pull requests carefully. That's why we're developing Tweety, an AI reviewer for pull requests, and experimenting with risk-based approvals modelled on Anthropic. The idea: not every change needs the same attention. How closely a person looks should depend on the risk of the change.

Animated by Claude Opus 5.5, the AI model that supports us on Exodus. Music: Suno.

I see this above all as an opportunity. Developers can spend their time where it makes the most difference. There is a risk all the same: if the AI writes most of the code, our ability to understand code in detail declines. In return, we can invest more time in requirements engineering, in other words in the question of what should actually be built.

How we deal with these opportunities and risks isn't something each team works out on its own. In our AI Coding Guild, developers and product owners have been sharing their experience with AI across teams since 2023.

What's next

Around 70 per cent of the stories are live. But the biggest chunk is still ahead: the data itself. 7.6 billion records have to move from Devinite to the new database, while everything keeps running.

That's why I can't say yet when Exodus will be finished. But one thing I'm sure of: when writing code is no longer the bottleneck, what happens before and after matters all the more. Preparing well what we want to build, and checking carefully what has been built.

¹ At around 128 GB, this table is the fourth largest of the 1,358 tables in Devinite.

3 people like this article


User Avatar
User Avatar

As the first full-time developer at Digitec Galaxus, I later was team leader of Goldfinger and Bender! and head of the architecture guild (the A-Team). Currently I am, together with our Domain Architects, only guardian of the ivory tower. As enabling team we permanently refine the target architecture, work closely with the development teams and form the Architecture Review Board.
Because of my unimaginable build breaks and piratical deploys on Friday afternoon they also call me 🅷🅰🅲🅺🅴🆁🅼🅰🅽.
 


Tech
Follow topics and stay updated on your areas of interest

Software
Follow topics and stay updated on your areas of interest

Behind the scenes

News about features in our shop, information from marketing and logistics, and much more.

Show all

These articles might also interest you

Show all

These articles might also interest you

Comments

Avatar