Ringier Slovakia Communities s.r.o. operates as a technology company managing online community platforms, including Pokec—one of Slovakia’s most active social networks. Founded as part of the Ringier Group, the company combines media operations with technological innovation, serving a diverse user base across different demographic segments in Slovakia.
The challenge: When content moderation can’t keep up
Messages on Pokec contained a huge amount of hate speech and inappropriate content. Their manual moderation system simply couldn’t handle the flood of user-generated posts—everything from elementary school slang to university-level discourse. The linguistic complexity, slang, and user creativity made automated solutions seemimpossible.
The team needed a smart system that could automatically approve safe content, reject clearly inappropriate posts, and flag borderline cases for human review. Speed was crucial, but so was accuracy. And there was one non- negotiable requirement: all user data had to stay on-premises to ensure privacy and comply with regulations.
What we set out to achieve
When we started our mentorship in November 2024, the goals were ambitious but clear:
- Build an AI-powered content moderation system using Large Language Models (LLMs) by summer 2025
- Dramatically improve moderation speed and accuracy
- Free up human moderators to handle complex cases requiring nuanced judgment
- Create a scalable solution that could work across other Ringier platforms and potentially become a commercial product
The team understood AI’s potential but needed guidance on choosing the right models, preparing data, optimising thresholds, and navigating implementation complexities.
The team brought a practical, hands-on approach from the start. Their willingness to experiment and iterate quickly made the technical transition smooth. Rather than needing a complete mindset shift, they naturally adapted to AI development’s iterative nature, building on their existing problem-solving skills.
Building technical expertise
Throughout our work together, we focused on developing practical skills:
- Data preparation: The team manually labeled 1,000 messages as approved, pending, or rejected—foundationalwork for training and testing models
- Model experimentation: We tested ChatGPT, Claude, and Gemini, comparing performance and costs to find what worked best for Pokec’s specific needs
- Threshold optimisation: We developed a three-tier system with carefully calibrated thresholds: auto-approve for clearly safe content, pending review (threshold 0.4), and auto-reject (threshold 0.9)
- Advanced techniques: By summer, we implemented RAG (Retrieval-Augmented Generation) with vectorisation and cosine similarity, achieving lightning-fast 60-millisecond response times
Overcoming key obstacles
Challenge 1: Language Complexity
Pokec users range from teenagers to professionals, creating huge variation in language sophistication. We solved this by fine-tuning smaller models and implementing RAG technology to better understand context across different communication styles.
Challenge 2: Finding the Right Balance
Setting thresholds was tricky. Too strict, and legitimate content gets blocked. Too lenient, and hate speech slips through. Through months of iterative testing (May through August 2025), we found the sweet spot that minimised pending reviews while maintaining accuracy.
Challenge 3: Speed Matters
Early versions worked but were too slow for real-time use. Through RAG implementation and model optimisation, we got processing delays down to just 60 milliseconds—fast enough for production.
Challenge 4: Going Multilingual
By August, the team recognised an opportunity to extend beyond Slovak. They added support for English, German, and Polish, and by November 2025, implemented Small Language Models (SLMs) from Hugging Face for language identification, making the system truly multilingual.
What we achieved
Technical Wins
- 60ms response time: Real-time content moderation in production
- Three-tier system: Auto-approve, auto-reject, or flag for human review
- RAG implementation: Vectorisation with cosine similarity for accuracy
- Multilingual: Slovak, English, German, Polish via SLM integration
- Custom models: Fine-tuned for Pokec’s specific context
Business Impact
- 50% less moderation hours: From 24 to 12 hours daily (freed up 1.5 FTE)
- 30% higher user satisfaction: Better content safety and platform reputation
- New revenue potential: Scalable solution for other platforms and external clients
Looking ahead
The success of this AI-powered moderation system has opened new doors for Ringier. The team continues refining the models, expanding language support, and exploring opportunities to deploy the solution across other Ringier platforms and potentially as a commercial product for external clients.
This case demonstrates how strategic mentorship combined with technical innovation can transform operational challenges into competitive advantages, ultimately contributing to the long-term resilience and sustainability of digital media platforms.
In their own words
Thanks to this mentorship, we are now able to filter out hate speech in milliseconds, and we are integrating AI into many more processes across our company. We also exceeded our goals and developed this solution for other languages in Central Europe—an important achievement since hate speech and inappropriate content continue togrow, and we are able to fight it efficiently. – Katarina Mikundova, Data Minding team