Four agents. One private stack.
Every product below runs on customer-owned GPUs against customer-owned data. They share one inference layer, one orchestration runtime, and one deployment model.
Forge
Design to fabrication. Plain-language part requests become validated, machine-ready files. A local model emits a constrained specification; deterministic code checks it against real machine limits and writes kerf-compensated output on separate cut and bend layers.
Failures return to the model with the offending value and the limit, driving an automated repair loop. Nothing unvalidated is written to disk.
Studio
Private generative media. Image and video generation on customer hardware, using brand-specific adapters trained on the customer’s own assets. Node-graph pipelines are driven programmatically, so a campaign variant set is a job rather than a session.
Built for teams working with unreleased product, licensed talent, or anything else that legally cannot be uploaded to a hosted service.
Intelligence
Quoting and capacity. The agent reads an incoming RFQ, resolves material against current stock and vendor pricing, checks machine queues for realistic capacity, and returns a costed quote with a defensible lead time.
It works from your constraints, not generic benchmarks: your stock, your queues, your vendors, your margins.
Checker
Engineering review. Named for the checker — the person whose signature sits in the title block of every released drawing. It ingests specifications, drawings, and standards held internally, then answers questions about tolerance stacks, wall thickness, draft, and cost drivers, citing the source document for every finding.
Retrieval runs against a local index. The corpus never leaves, and neither does the query.
What every product inherits.
Air-gappable
Full function with no outbound connection. Nothing is transmitted, logged off-site, or used for training.
Fixed cost
Capital hardware rather than usage billing. Unit economics improve as volume grows instead of degrading.
Model-portable
Open-weight models swap without re-architecting. No single provider can deprecate your workflow.