Sovereign AI. Your data, your GPU.
An OpenAI-compatible endpoint that runs entirely inside your network. Embeddings, semantic search, and retrieval-augmented generation over your own data — without it ever crossing the boundary to a third-party LLM provider.
The AI surface, kept inside the perimeter.
OpenAI-compatible API
/v1/embeddings and /v1/search drop in wherever your code expects OpenAI. The migration is the base URL, the API key, and nothing else.
ONNX runtime on DirectML
Models export as ONNX, execute on DirectML — meaning any DirectX-12 GPU in the customer estate works. NVIDIA, AMD, Intel Arc, integrated. No CUDA lock-in.
Coordinator + agent topology
ASP.NET Core coordinator schedules work; a WPF agent on each GPU host runs inference. The coordinator never sees raw data — agents pull, embed, return vectors to your application.
Air-gap deployable
No internet egress required at inference time. Model weights distribute via signed bundles. Licence verification has a 24-hour offline grace window — built for SCIF and defence environments.
Multiple model families
Ships with all-MiniLM-L6-v2 (general embeddings), Phi-mini (small LLM), BGE (retrieval-tuned). Bring your own ONNX model and register it as a new model name.
Bound to SkyeConnex storage
The /v1/search endpoint queries the SkyeConnex content index. RAG over regulated data without copying it — embeddings stay on the same side of the perimeter as the source.
Two services. One sovereign envelope.
SkyeGXU coordinator
ASP.NET Core service running inside the customer cluster. Surfaces the OpenAI-compatible HTTP API. Schedules embedding and search work across registered agents. Maintains the vector index in the SkyeConnex storage layer — encrypted and RAID-scattered like every other file.
Coordinator deploys as a container or a Windows Service. Authenticates inbound requests through the SkyeConnex IAM. Surfaces Prometheus metrics for queue depth, agent health, and per-model latency.
SkyeGXU agent
WPF application running on the GPU host. Polls the coordinator for work, executes ONNX inference via DirectML, returns embedding vectors or search results. Stateless — an agent can be killed and replaced without losing in-flight work.
Agents register themselves with the coordinator at startup, advertising which models they have loaded. The coordinator routes work based on model availability and current load. Multiple GPUs per host, multiple hosts per cluster.
Why sovereign AI is no longer optional
Through 2024-2026 the regulatory posture around sending regulated data to third-party LLM providers has tightened sharply. Healthcare data handling agreements increasingly call out OpenAI and Anthropic explicitly as out-of-scope third parties. EU AI Act high-risk system requirements implicate the embedding pipeline alongside the model. Defence and intelligence customers have always had this as a hard requirement; what has changed is that healthcare, financial services, and legal organisations now have it too.
What runs on SkyeGXU
The shipping model set covers the most common embedding and retrieval workloads:
- all-MiniLM-L6-v2 — 384-dimensional sentence embeddings, fast on CPU and accelerated on GPU. Good general-purpose retrieval.
- BGE-large-en (BAAI General Embedding) — higher-quality retrieval embeddings tuned for English. Slower, better recall on regulated-corpus retrieval.
- Phi-mini — small LLM for retrieval-augmented generation. The RAG response generator that pairs with the embedding pipeline above.
Custom ONNX models can be registered: drop the model file, advertise its dimensions and tokenizer, and the coordinator routes /v1/embeddings calls naming that model to the registered agent.
Worked example: clinical RAG without OpenAI exposure
A hospital wants Barrista (the SkyeConnex AI assistant) to answer questions over patient records without records crossing the perimeter. The deployment:
- SkyeConnex storage layer with PHIPA-aligned residency policy (Ontario providers only).
- SkyeGXU coordinator deployed in the hospital's VPC. WPF agents on two clinical-AI workstations with consumer-grade NVIDIA RTX GPUs.
- Patient-record ingest writes documents to SkyeConnex storage. The coordinator embeds them via the agents using BGE-large-en. The embeddings live alongside the source.
- Clinician asks Barrista a question. The query embeds locally on a GPU agent. Top-k candidate records are surfaced. Phi-mini composes the answer, citing the source records.
No patient record leaves the hospital network. The data-handling agreement language about OpenAI is satisfied because OpenAI is not in the architecture.
Run AI on your own data.
Book a session. We'll demonstrate the embedding pipeline over a sample regulated corpus — entirely on the customer side of the perimeter.