DigisixDigisixBusiness Analysis
dorothee@digisix.be
Case · Personal project · in production

LudoExplorer, a personal project built end to end

A personal project that turns public board game data into a recommendation site, from data analysis through to production.

Role
From analysis to production
Since
2025
Languages
FR · NL · EN
Status
In production
Visit ludoexplorer.com →

The challenge

LudoExplorer is a personal project, built entirely from scratch, that turns a public dataset about board games into a live recommendation experience. It is also a concrete demonstration of the four areas I work on for clients: business and data analysis, automation and applied artificial intelligence.

Board game enthusiasts have thousands of titles to choose from, but few simple ways to find a game that genuinely fits their preferences, group size, language or willingness to read a lot of rules. Public data exists on the subject, but it is large, messy and multilingual and cannot simply be used as it is.

The friction points at the start

  • Friction 01

    Thousands of titles

    Few simple ways to find a game that fits your real criteria.

  • Friction 02

    Raw data

    Large, messy and multilingual, not usable as it is.

  • Friction 03

    Insider vocabulary

    Hard to find a starting point without knowing the hobby's terms.

The approach

Before a single line of recommendation code was written, the real work was in the data. Making sense of a large and inconsistent public export, with games described in several languages, inconsistent categories, missing fields and duplicates, and deciding what was actually worth keeping. Building a classification that reflects how people really search for a game, not simply how the raw data happened to be labelled. And defining what a good recommendation actually means: a combination of similarity, popularity and fit with player count, complexity and language.

Once that foundation was in place, the data was enriched. Short, purpose-written descriptions replaced the often dry and inconsistent source text, in all three languages of the site.

The automation layer keeps the site running without manual intervention, with updates, backups and health monitoring on a production server.

Artificial intelligence is used where it genuinely adds value: a personalised recommendation engine, more flexible interpretation of search queries, and a personality quiz that gives people a starting point without requiring them to already know the technical vocabulary of the hobby.

  • Deep learning
  • Applied AI
  • Personal project
  • Machine learning

What was delivered

  • Structured data

    Cleaned, classified the way people really search for a game, enriched in three languages.

  • Recommendations

    A mix of similarity, popularity and fit with player count, complexity and language.

  • A site that runs itself

    Automatic updates, backups and health monitoring in production.

The result

A public website with search, filters, detailed game pages, personalised recommendations, a tool for comparing two games and a personality quiz. Full support in French, Dutch and English for search, recommendations and content, not just labels. An internal dashboard that tracks real usage and errors: with little traffic so far, it mostly helps to catch problems early. A production deployment that runs unattended, with automatic updates, backups and health monitoring.

What the project made possible

  • A complete public site

    Search, filters, game pages, recommendations, comparison and a personality quiz.

  • Three languages

    French, Dutch and English, for the content and not just the labels.

  • Problems caught early

    An internal dashboard tracks usage and errors. With little traffic so far, it mostly shows what is broken.

Behind the scenes

I wrote down how the project was built, mistakes included, in ten articles on Medium. The lessons that also apply to a client project are in this article.

Key takeaways

LudoExplorer is small enough to explain on a single page, yet it touches every layer of my work.

  1. Filter the data against the need

    About 125,000 entries in the source, about 90,000 kept in the catalogue.

  2. Ask instead of guessing

    Search has two modes, by name or by idea, because no rule could tell them apart reliably.

  3. Count people, not page loads

    259,106 sessions in four weeks, 136 of them with a real action.

  4. Measure before believing

    A convincing fix would have rewritten 5,441 correct search keys. I checked first.

Is one of your processes eating up your time? The Process Scan gives you a first reading in 5 minutes, free. Take the Process Scan →

Data, AI, automation and user experience can come together without losing sight of the original problem

If your project combines several of these areas, I can help connect the business need, the data and the practical solution.

Discuss my project →dorothee@digisix.be