Loading
Loading
Controlled, evidence-based research that advances Romanian and European AI capability — with clear stage gates, published evaluation results and rigorous data governance.
NOVACORE AI approaches research as a structured programme, not a speculative activity. The research division focuses on a small number of carefully scoped priorities — language models, Romanian-language AI, evaluation benchmarks and applied research with measurable outcomes — and progresses through disciplined stage gates before committing resources to larger efforts. Every research track begins with a clear hypothesis, a defined evaluation methodology and a commitment to publish evidence — whether the results confirm or refute the hypothesis. Models and datasets are released only when they are properly documented, evaluated and governed. The programme does not pursue artificial general intelligence or speculative long-term research. It focuses on practical contributions to European AI sovereignty: models that work well in Romanian, evaluation frameworks that measure real-world performance, and datasets that are properly licensed, documented and curated. This discipline reflects NOVACORE's broader philosophy: build only what can be evidenced, report only what has been verified, and invest only where capability, data and infrastructure are sufficient to deliver credible results.
Research into language model architectures, training methodology, fine-tuning techniques and evaluation. Focus on smaller, efficient models that perform well on specific tasks — not indiscriminate scaling.
Developing AI capabilities for the Romanian language — datasets, evaluation frameworks, fine-tuned models and benchmarks. Building genuine Romanian-language AI, not just translated English models.
Performance benchmarks, evaluation frameworks and comparison methodologies. Transparent measurement of model capability — what models can actually do, not what marketing claims suggest.
Focus on business applications, Romanian use cases, privacy-preserving deployment and rigorous evaluation. Build capability and evidence before investing in model development.
Domain adaptation and Romanian-language optimisation with measurable evaluation results. Publish fine-tuning recipes, evaluation data and performance comparisons.
Licensed Romanian, legal, technical and selected regional text only. All training data must be properly licensed, documented and governed. No indiscriminate web scraping.
Only when demand, data, staff and GPU capacity are in place. Models for specific domains — legal, medical, technical — with validated performance and clear use cases.
Formal feasibility review before any large-scale training commitment. Capital requirements, compute budget, data availability, evaluation framework and publication plan must all be in place.
NOVACORE AI does not announce research results before they are verified. Model releases are accompanied by model cards, evaluation benchmarks and data-provenance documentation. Training data is properly licensed and documented — no indiscriminate scraping of copyrighted material. Research partnerships require formal agreements with clear IP, publication and data-governance terms. Every claim about model capability published by NOVACORE must be supported by reproducible evidence. This discipline may slow publication cadence relative to some industry norms, but it ensures that what we publish can be trusted.
Secure AI and high-performance computing for enterprises, governments and research.