Agentic engineering with deterministic acceptance gates.
Forge is a research platform for reducing iteration time across specifications, hardware, verification, firmware, and kernels—without trusting generated artifacts by default.
Research platform; human engineering judgment remains responsible
Specialized assistance across one engineering trace.
Candidate roles cover specification analysis, workload characterization, design-space exploration, RTL and DV scaffolding, firmware, kernels, documentation, and regression triage.
Model confidence is never the acceptance criterion.
Artifacts move through deterministic compilation, lint, reference comparison, randomized tests, assertions, coverage, synthesis or performance regression, and review.
Generate scaffolding, tests, and reviewable changes—not unquestioned hardware.
Assistance may produce microarchitecture scaffolds, test plans, cocotb tests, assertions, and triage summaries. Independent tools must verify every accepted result.
Target constraints and differential tests surround generation.
A declarative kernel specification can lead to RVV or matrix-targeted code, but compilation, randomized shapes, edge cases, tolerance checks, and profiling decide whether it advances.
Accepted output becomes a versioned artifact with evidence.
Inputs, tool versions, generated changes, deterministic checks, failures, and review decisions are intended to remain traceable.
Architecture and acceptance remain engineering responsibilities.
Agents may accelerate analysis and implementation, but human engineers own requirements, architecture, exceptions, and final sign-off.
The system is a prototype direction, not autonomous hardware development.
No claim is made that generated RTL, firmware, tests, or kernels are correct without independent verification.