Product · Phase 1 operational
Forge
Describe a part in plain language. Forge produces a specification, checks it against your machines, repairs what fails, and writes a cut-ready file. Every step runs on your own GPUs.
StageDeveloping
RuntimeCUDA accelerated
DeploymentOn-premises
OutputDXF · JSON
# plain language in, machine-ready file out, no external call $ python run.py "aluminium mounting plate 180x120mm, four M5 holes near the corners, two M4 holes on 40mm centres, cable slot above them" -- attempt 1 -------------------------------- DFM: 2 error(s), 0 warning(s) [FAIL] EDGE_DISTANCE: mount-BL clearance 3.20mm < minimum 4.00mm [FAIL] MIN_WEB: web between device-L and cable-pass is 1.40mm < minimum 2.00mm -- attempt 2 -------------------------------- DFM: clean — no findings. blank : 180.0 x 120.0 mm features : 6 holes, 1 slot, 0 bends DXF -> out/APEX-BRKT-01.dxf
The repair loop is the point. A model producing an unmanufacturable part is expected. The system catching it by rule — with the offending value and the limit it broke — is what makes the output trustworthy.
Pipeline
Five stages. Only one is probabilistic.
STAGE 1RequestPlain language, typed or spoken. No CAD skill assumed.
STAGE 2SpecifyA local model emits a constrained JSON spec — never geometry.
STAGE 3ValidateDeterministic DFM rules check the spec against real machine limits.
STAGE 4RepairFailures return with values and limits. The model revises.
STAGE 5EmitKerf-compensated DXF on separate cut and bend layers.
Design decision
The model never emits geometry.
Language models are unreliable at coordinate arithmetic and file formats, and highly reliable at filling in a small schema from a description. Forge splits the work along exactly that line.
A hallucination therefore surfaces as an out-of-range parameter, which the validator catches by rule — not as malformed geometry that reaches a machine. Kerf compensation and bend arithmetic happen in deterministic code for the same reason.
- RulesMinimum hole diameter versus material thickness
- Edge distance and web width between features
- Press-brake flange minimums and bend length
- Bed capacity and blank fit
- Feature proximity to bend lines
- ResultNothing unvalidated is written to disk. A failing spec exits without producing a file.
Deployment
What installing it looks like.
- Machine surveyWe measure your actual limits — bed capacity, kerf, minimum features, brake tonnage and flange minimums — and encode them. The rules are yours, not defaults.
- Hardware specificationSized to workload and budget. A single workstation-class GPU is enough for most shops; larger deployments scale across nodes.
- Model deploymentOpen-weight models served locally. We benchmark two or three against your part vocabulary and pick on measured behaviour.
- IntegrationOutput lands where your CAM software already looks. No new workflow to learn if the existing one works.
- HandoverYour team runs it. The system keeps working whether or not we are still involved.
Where it fits
Built for shops where the drawing cannot leave.
Sector
Defence & aerospace supply
Machine shops and tier suppliers under controlled-information requirements, where a cloud endpoint is not an option at any price.
Sector
Contract manufacturing
Job shops quoting and producing against customer geometry covered by NDA, needing faster turnaround without exporting client data.
Sector
Industrial OEMs
Producers digitising quoting, planning, and engineering knowledge held in legacy systems and long-tenured heads.
Current status
Phase 1 is operational on production hardware: structured generation, eleven DFM rules, the automated repair loop, and kerf-compensated DXF output. Phase 2 extends to tool-calling orchestration with retrieval over private document stores. Forge is not yet generally available — we work with a small number of early partners.
Want to see it run against your machines?
We will encode your equipment limits and generate a real part.