The Lab
The private AI cluster pattern — running, not theoretical.
This is a reference implementation of the architecture Tradewinds offers to clients, built and operated in a lab environment. It is not a client deployment — it's the proof that the pattern works before it's recommended to anyone.
The architecture
The reference build implements the pattern a CUI-handling contractor needs when commercial AI tools are off the table: a self-contained inference cluster — GPU compute serving open-weight language models behind an internal gateway — inside a segmented network with no path to the public cloud. Users get chat, retrieval, and API access to capable models; the data never crosses the boundary.
Design decisions
Local inference
Open-weight models served entirely on owned hardware. Prompts, outputs, and embeddings never leave the environment — the capability lives where the data lives.
Zero cloud egress
The inference segment has no outbound path to commercial AI services. Egress is denied by default and model updates arrive through a controlled, auditable process.
Model routing
A routing layer matches each request to the right model for the job — larger models for complex reasoning, smaller ones for fast, cheap tasks — without callers needing to know the difference.
Firewalled, segmented access
Inference services sit in their own network segment behind explicit allow rules. Clients reach a gateway, not the machines — the same posture an enclave deployment requires.
What it delivers
In general terms: interactive chat at conversational speed for a small team, batch and retrieval workloads that run overnight without touching an external API, and the ability to swap or add models as open-weight releases improve — all at a fixed, predictable hardware cost instead of per-token billing. The same pattern scales from a single node to a multi-node enclave deployment; the design decisions stay identical, only the hardware count changes.