SouthByte

SouthByte advises companies that already use AI and need to know what that means legally and operationally. The work is hands-on rather than paper-only: model selection, on-premise deployment, risk classification under the EU AI Act, and the governance around it.

This org exists to host what can be shared publicly from that work. Right now that is mainly the evaluation results linked below — the models themselves are third-party, so there is nothing to re-upload.

Public artifacts

Model evaluations — results.southbyte.de

Own test rig for on-premise operation, run on an NVIDIA DGX Spark (GB10). Current run: 45 language models served under vLLM, 5 guard models, 6 image models and 16 TTS models, scored on answer quality, German-language capability, bias, code and throughput. Machine-readable summary: summary.json.

The headline metrics are public. The security-relevant detail reports are deliberately not — publishing a ranked list of which guard models fail on which attack class helps the wrong people. If you work in AI safety and need them for research, get in touch.

German PII recognizers for Presidio — data-privacy-stack/presidio

Contributed the DE_* recognizers for German personally identifiable data, including checksum validation implemented against primary sources. Presidio is the widely used open-source framework for detecting and anonymizing PII — a building block most German AI projects need sooner or later.

What you will not find here

No fine-tunes of other people's models re-uploaded under a new name, no benchmark claims without a reproducible setup, and no "AI-powered" anything. If something is published here, the method is documented and the numbers can be checked.

Contact

kontakt@southbyte.de · southbyte.de · LinkedIn · GitHub

Deutsch

AI Governance und IT-Beratung für den Mittelstand. EU AI Act, ISO/IEC 42001, On-Prem-Betrieb von Sprachmodellen. Großraum München.

Diese Organisation hostet, was aus der Beratungsarbeit öffentlich geteilt werden kann. Derzeit sind das vor allem die Modell-Evaluationen unter results.southbyte.de: 45 Sprachmodelle unter vLLM, 5 Guard-Modelle, 6 Bildmodelle und 16 TTS-Modelle auf einer NVIDIA DGX Spark, bewertet nach Antwortqualität, Deutschfähigkeit, Bias, Code und Durchsatz.

Die Kennzahlen sind veröffentlicht, die sicherheitsrelevanten Detailberichte bewusst nicht. Dazu kommen die deutschen PII-Erkennungsregeln (DE_*) für Presidio, inklusive Prüfziffernvalidierung nach Primärquellen.

Kontakt: kontakt@southbyte.de