Why not just a big model
A demo describes a CDC fix.
A production system proves it.
An LLM can talk about clock-domain crossings. It cannot reliably run Verilator with the right flags, parse the real netlist, or prove RTL equals gate — and it invents structure. Scripts are deterministic but brittle and silent. Five properties separate a system you can sign off with from a system you can only be impressed by.
Architecture
Four tiers, and the request only travels one way
Routing is separate from planning, planning is separate from execution, and everything is grounded in a parsed model of your design rather than a description of it.
Request lifecycle
Domains
Seven Leads, one workspace
Each Lead plans within its domain and threads its collateral to the next — a CDC waiver set reused downstream, a synthesised netlist handed to DFT, a waveform routed back into regression debug.
What is different
The layer above your existing tools
Each vendor's AI is locked inside that vendor's tools. This is the orchestration and knowledge layer that drives all of them together — and that your team can extend.
Surfaces
Same SDK, six ways in
A browser for the plan card and Step Inspector, a terminal for the people who live there, a laptop when there is no farm, and MCP when something else should drive it.