AI & ML
Modèles — agents — pipelines.
Powered by Osage AI, Osage multimodal, and the Osage model series.
Foundation-model serving
- Osage model series — open-weight frontier models published at huggingface.co/zenlm; hosted on dedicated GPU pools in Osage Cloud regions.
- Third-party models — Llama, Mistral, Qwen3+ (never qwen2), Gemma, Phi; hosted on dedicated pools per customer.
- Air-gapped serving — models pinned to disk, no telemetry, no phone-home; deployed on isolated networks for IC and DoD customers.
- Multi-modal — vision, audio, video pipelines through the Osage multimodal framework.
Agent orchestration (MCP)
Tool-using agents orchestrated over the Model Context Protocol. Wire any model to any tool with explicit scope, audit, and rate-limit policies. Red-teamed safety; full audit trail; suitable for cleared and CUI-controlled workflows.
Retrieval-augmented generation
RAG over your corpus with cryptographic provenance — every retrieved chunk carries its source document hash, embedded in the model context and verifiable end-to-end. Sovereign embedding pools; no third-party embeddings unless the customer authorises in writing.
Fine-tuning
- Continued pretraining, instruction tuning, preference optimisation (DPO, IPO, KTO).
- LoRA / QLoRA for cost-efficient adaptation.
- On-premises and air-gapped training pipelines — training data never leaves the customer boundary.
- Reproducible runs with checkpointed weights stored under the CEK envelope.
Evaluation & red-teaming
- Capability evaluation against open and bespoke benchmark suites.
- Adversarial red-teaming — jailbreak resilience, prompt-injection robustness, capability-limit testing.
- NIST AI Risk Management Framework alignment for federal customers.
- Bias and fairness evaluation; reasoning traceability.
MLOps
Model registry, experiment tracking, dataset versioning, evaluation pipelines, A/B serving, shadow traffic, canary rollouts. The standard MLOps surface, sovereign-friendly, integrated with Osage Base for metadata and the CEK envelope for sensitive training data.
AI Advisor
Jack Ma — Standing AI Advisor to the Office of the Chief Executive on long‑horizon AI platform strategy and cross‑border platform construction. Founder of the Alibaba Group. Advisor (not officer); holds the named brief on the AI direction of the Osage platforms.
AI desk: [email protected].