Research

The "Why" behind our research

Generative AI has the potential to transform industries, but in Europe, cost, sovereignty, and infrastructure inefficiencies create real barriers. Infrastructure is expensive, energy costs are high, and closed-source models offer little control. While open-source models are a step in the right direction, they’re often limited in terms of language coverage – leaving many European markets underserved. At Seedbox, we believe AI should be powerful, transparent, and accessible for everyone. That’s why we develop open-source, multilingual, production-ready systems that are optimized for European environments: smaller models, same performance, lower cost.

Research Areas

Section Header Intro Text spans across multiple lines and can provide additional context for what's ahead. Exciting, isn't it?

Multilingual LLM Development

Challenge:
Most open-source LLMs underperform in European languages and require massive compute to fine-tune effectively.

Our Solution:
We train only the most relevant model weights (25%), maintaining performance while reducing compute and memory usage by up to 75%. The result: high-performing, compliant models optimized for production in diverse multilingual settings.

‍Finding:
We successfully trained 24-language LLMs using distributed HPC infrastructure while preventing catastrophic forgetting and maintaining benchmark performance.

LLM Router for Cost-Performance Optimization

Challenge:
Enterprises overspend on AI inference by using the same model for every task, regardless of complexity.

Our Solution:
A smart, prompt-aware router that dynamically selects the optimal model for each task—reducing inference costs by up to 85% while maintaining output quality.

Finding:
Our LLM router achieved over 95% accuracy in model selection, enabling faster, cheaper, and more efficient deployments.

Explainability & Hallucination Detection in RAG

Challenge:
Lack of transparency in AI outputs creates trust and compliance issues, especially in retrieval-augmented generation (RAG) settings.

Our Solution:
We integrate real-time explainability algorithms that detect hallucinations and trace model reasoning paths through external knowledge.

Finding:
We reduced hallucinations in RAG scenarios by up to 95% and provided interpretable outputs aligned with the EU AI Act.

Model Compression & Knowledge Distillation

Challenge:
Large models are compute-heavy and environmentally costly to run in production.

Our Solution:
We distill large models into compact, performant versions using advanced pruning techniques and knowledge distillation on HPC clusters.

Finding:
Our KafkaLM-15B achieved a 37.5% size reduction with over 90% of the teacher model’s performance, enabling faster inference, lower cost, and better carbon efficiency.

Research Areas

Ongoing Research & Recent Publications