Building the Data and ML backbone at Adevinta

A project that transformed AI and Data capabilities in Adevinta

Orchard was Adevinta’s post-eBay integration and synergy program. My work was focused on the Data Foundations and ML workstream: consolidating fragmented capabilities and building new platforms to accelerate ML, data and software engineering teams while delivering measurable synergies.

As part of this program, we built or scaled:

  1. Machine Learning Platform
  2. Computer Vision Platform
  3. ML Enablement team
  4. Data Catalogue
  5. Experimentation Platform
  6. Curated Dataset Platform

My role started with identifying the need, building the business cases, hiring the new teams, and overseeing the development of the AI-related platforms and capabilities (first three in the list above). Later, I was also responsible for ensuring that all the platforms delivered the expected synergies. With a one-off investment of €13M over two years, the platforms generated €14M in measured annual synergies.

This project transformed how Adevinta and its marketplaces built and maintained data and ML products, consolidating fragmented team-level stacks into state-of-the-art platforms used by hundreds of engineers and significantly improving time-to-market KPIs.

Learnings

Building and scaling ML platforms and teams taught me many lessons. One of the most important was the need to build credible business cases early. Securing alignment and planning the work ahead may sound simple, but it required significant effort, especially given the level of investment involved.

Once the decision was made, the challenge shifted to hiring, building the teams, and delivering. At Adevinta, many fragmented solutions already existed because engineering teams still needed to build ML and computer vision products. However, most of those solutions were subpar, inconsistent, or solved only part of the problem.

Two strategic decisions proved especially important. First, we built the platforms as products, not just as internal infrastructure. Second, we did not force teams to adopt them. Each platform had product management ownership, and the teams stayed very close to user needs. We involved practitioners directly in shaping what the platforms should become and created an active community around them to capture needs, test technologies, share progress, and demonstrate what was possible.

By focusing on delivering genuinely useful solutions and staying close to the teams, adoption grew quickly. Eventually, our platforms were used by 90% of the teams building data and ML solutions.

We saw many examples of the impact: teams built ML and computer vision products in days rather than weeks, accelerated experimentation, and translated that speed into better experiences for millions of users across Adevinta’s marketplaces.