Not a chat layer over a solver. A learned model of the fields, trained on physics we generate, and fast enough that exploring a design space becomes a search instead of a handful of guesses.
Every label comes from a solver we run ourselves, drawn to cover the designs people actually build.
A field at a point, a response between ports, a figure of merit over a band. One model, whatever the question needs.
Milliseconds per candidate where a solve is minutes to hours. That is what turns exploration from intuition into a search.
An agent takes the spec, proposes changes across the whole system, scores them on the model, and sends only the survivors to the solver. What comes back is a ranked set of proposals with physics behind every number.
You stop searching with the solver and start checking with it.
A full-wave solve is minutes to hours, so exploration gets rationed to a handful of candidates picked by intuition. Scoring thousands changes which designs you consider at all — and nothing ships on a prediction alone.
00:00 → read_system die · package · board · VRM
00:01 → propose decap, vias, cap regions, crossover
00:02 → model_sweep 2,048 variants
00:06 package-only fix misses 10 MHz — dropped
00:07 → promote top 5 → solver
01:38 3 proposals ranked · one change per team
An agent reads the spec and the system as built, then proposes variants at every level — die, package, board, regulator — not just the layer whose file you happen to own.
Thousands of variants are scored in the time a solver takes for one. Dead ends are dropped early, and infeasible specs are called before anyone spends a solve on them.
Survivors are re-run through the real toolchain, and the numbers in a proposal are the solver’s, never the model’s. The agent can edit the design, never the checks.
The same model and the same agents, pointed at four problems. Each takes a spec you already write down and returns a verdict saying which lines are met, on what evidence.
Planar antennas with the laminate free to change, since it is often the substrate that decides whether a spec is reachable. Bring the band, match, gain and envelope; get a geometry and feed with a verdict on every spec line, and impossible specs called early.
A rail from regulator to die, closed against a target impedance and a droop budget rather than a rule of thumb about decap counts. Bring the stack and the target; get ranked proposals naming what the die, package and board each change, with the solver behind every number.
High-speed channels end to end: the stackup, the via transitions and the routing that decide whether the eye is open at the receiver. Bring the data rate and the stackup you can build; get a routed channel and a verdict on loss, crosstalk and eye margin.
Distributed filters, couplers and matching networks synthesised against a response instead of tuned by hand. Bring the passband, rejection and the area you can spend; get a layout and a verdict on every line of the response, including where laminate tolerance eats the margin.
Thirty minutes: the spec you are trying to close, and the constraint that keeps biting. We run it and show you what comes back, including where it falls short.