YourAIrunsonyourinfrastructure.Yourdataneverleavesit.

Each model has different strengths, like different medical specialists. GPT-4o (OpenAI) excels at versatile tasks and complex reasoning. Claude (Anthropic) is known for nuanced responses and caution on sensitive topics — ideal for legal and compliance. Mistral is a lighter European model, optimized for French, less computationally expensive. The good news: you don't need to choose manually — SkaLean's LLM router automatically selects the optimal model based on task type, language, compliance constraints, and cost target.

GPU inference is the computation performed to generate each AI response. Normally, with ChatGPT, this computation happens on OpenAI's servers in the United States, subject to the US Cloud Act — your data transits to a foreign country. Sovereign GPU means SkaLean performs this computation on GPUs physically located in your country. For a Quebec medical office (Bill 25), a law firm (professional secrecy), or a financial institution (OSFI), this is a legal requirement. If an audit asks "where is your data processed?", the answer is "on a physical server in your city."

The LLM router evaluates each request on 4 criteria: (1) Presence of personal data — if the request contains sensitive data, only sovereign models are authorized, (2) HIPAA compliance — if the account is in HIPAA mode, US cloud models are excluded, (3) Client preference — if you have defined a preferred model for a specific use case, it takes priority, (4) Intelligent degradation — if the preferred model is unavailable, automatic failover to the best available alternative without service interruption. This system eliminates vendor lock-in.

Yes, available on Enterprise plans. LoRA fine-tuning (Low-Rank Adaptation) adjusts a base model (Llama 3, Mistral) on your specific data in 2 to 5 days. Use cases: dental office fine-tuning on RAMQ code nomenclature, accounting firm on specific Quebec tax regulations, veterinary clinic with its animal medical terminology. Result: 15-30% superior precision on your specific tasks. Most clients don't need it — the RAG pipeline suffices — but it's available if your vocabulary is highly specialized.

GPT-4o and Claude achieve very close performance in French and English (5-10% difference according to benchmarks). Mistral has been specifically optimized for French and often outperforms GPT-4 on French writing tasks. For Arabic, available models support Modern Standard Arabic (MSA) with good quality. Regional dialects (Darija, Levant, Gulf) are supported for simple conversational tasks. During your demonstration, SkaLean makes it easy to compare models side-by-side on your real use cases.

SkaLean manages updates in "blue-green" mode: the new version is tested in parallel for 72 hours before replacing the old one. If quality metrics regress on your use cases, the switch is automatically canceled. You are notified 7 days before any major update. For Enterprise plans, a "pinned model" (fixed version) can be configured to avoid any unplanned behavior change — unlike direct OpenAI/Anthropic APIs where an update can change your application overnight.